BitSquare/Codeman
2
0
Fork 0
mirror of https://github.com/Ark0N/Codeman.git synced 2026-10-03 22:19:42 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge master into PR #156 (remote tmux durability)

Browse Source 
Resolutions:
- session.ts: keep the extracted _buildRespawnPaneOptions() helper (COD-108)
  and add master's docker/owner fields to it
- tmux-manager.ts: docker branch first, then remote via buildRemoteSessionCommand
  (now an options object threading claudeMode/allowedTools into
  buildRemoteLaunchCommand, preserving the 6.3 multi-user permission downgrade)
- case-routes.ts: keep master's adminOnly helper; gate the new COD-105 discovery
  endpoint admin-only in multi-user mode (hosts are machine-level infra)
- settings-ui.js: union of remoteAutoReconnect + master's header-button defaults
- session-routes.ts: union of imports; session gets remote + owner

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
Codeman maintainer
2026-07-20 14:27:02 +02:00
parent 115ada1e9e d6917e3b21
commit 25c67f9415
88 changed files with 11862 additions and 616 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/docker-cases-plan.md
+433
View File
@@ -0,0 +1,433 @@
<!-- Design doc generated via ultracode multi-agent workflow (wf_e3a7498b-26f): 3 architecture proposals -> judge panel -> synthesis -> completeness critic. -->
# Docker Session Mode, Implementation Plan
## Decisions (locked 2026-07-19, by repo owner)
1. **Isolation posture**: CONVENIENT default (bind-mount host `~/.claude` etc. read-write so the existing login just works; network on; still hardened non-root + cap-drop + resource caps). SEALED profile (`mountCredentials:false` + `network:none`) is a per-case opt-in.
2. **Export**: offer BOTH full-image (`commit`+`save`+workspace tar) AND workspace-only, side by side, no default (ask each time).
3. **Base image**: BUILD LOCALLY on first use via `scripts/build-agent-image.mjs` from a repo `docker/agent.Dockerfile`. No registry required. (GHCR pull can be added later.)
4. **Hooks**: WIRE HOOKS NOW. Codeman scaffolds `.claude/settings.local.json` + CLAUDE.md into the linked host workspace dir (same as local cases), enabling in-container permission prompts, hook-idle detection, and the Claude Model picker.
Adopted defaults for the remaining open items (Section 10): resume-on-restart ON; container is per-CASE and shared by multiple sessions (killing one session only kills its in-container tmux session, never `docker stop` while siblings remain; stop/remove only on explicit teardown or case-delete); rootless caps = ship-with-warning (`capsEnforced` surfaced); remote docker daemon = local-first; podman = docker-first best-effort.
## Implementation status (branch `feat/docker-session-mode`)
DONE and END-TO-END VERIFIED against a real docker daemon (create host, link case, quick-start shell in a real container, workspace bind-mount round-trip, hook scaffolding, session-delete keeps the shared container up, case-delete `docker rm`s it):
- Phase 0-1: types (`DockerHost`/`DockerCase`/`SessionDocker`), `src/docker-hosts.ts` (storage, pure `buildDockerBaseArgs`/`buildDockerCreateArgs`, `containerApiUrl`, `hostGatewayAlias`, config-hash, credential-mount resolution, daemon probes), `DockerHostSchema`/`DockerCaseLinkSchema`. 26 unit tests.
- Phase 2: `tmux-manager` `buildDockerLaunchCommand` (image-check -> ensure -> start -> exec, resume-aware), `buildDockerKillCommand` (in-container tmux only, multi-session safe), stop/remove; wired into `createSession`/`respawnPane`/`killSession`. 14 unit tests.
- Phase 3: `Session` threading (`_docker`, toState, option builders, in-container cliVersion probe, `resolveMuxAttachCwd`), `server.ts` recovery round-trip.
- Phase 4: `case-routes` `/api/docker-hosts` CRUD + `/api/cases/docker-link` + listing + docker-unlink; `session-routes` `/api/quick-start` docker branch (rejects per-session config, probes availability + tmux, scaffolds hooks, seeds resume id).
- Phase 5 (partial): `docker/agent.Dockerfile` + `scripts/build-agent-image.mjs` (built + verified: node 22, tmux, claude/codex/gemini/opencode, arbitrary-uid HOME). Host-guard allowlists `host.docker.internal`/`host.containers.internal` for in-container hooks.
- Full CI green (3445 tests).
REMAINING:
- Phase 6: export / import (`docker commit` + `save | gzip` + workspace tar + manifest; `load` + quarantined re-tag), GC / boot reaper, disk-safety prechecks, drift-recreate route, SSE `docker:*` events. THE "move to a new machine" feature.
- Phase 7: frontend Create Case "Docker" tab + `linkDockerCase` + run wiring + case-picker labels + export/import UI.
- Phase 8: CLAUDE.md "Docker cases" Key Pattern + `docs/docker-cases.md` + COM.
- Deferred refinements: in-container model-picker via `settings.local.json`; live mid-run resume-id capture into `DockerCase.lastClaudeSessionId`; rootless/Desktop uid probe (currently a platform heuristic).
## 1. Goal & user stories
Add "Docker cases" to Codeman: a case can point at a container instead of a local or remote-SSH path, and any of the five CLI backends (`claude` / `shell` / `opencode` / `codex` / `gemini`) runs inside that container. It is modeled as a LOCATION OVERLAY on cases, exactly like the remote-SSH feature (COD-94/#145), never as a sixth `SessionMode`.
User stories:
- As the repo owner, I link a case to a per-project container so an autonomous Claude/Ralph run executes in a hardened sandbox (cap-drop, non-root, resource caps) instead of directly on my host, while keeping my existing OAuth login and transcript history working with zero extra setup.
- I set default, per-case-changeable container settings (image, network mode, memory/cpu/pids caps) at link time and edit them later, and edits actually take effect through a recreate-on-drift path (see Section 4).
- I reconnect after a Codeman restart and land back in the SAME running agent with the conversation intact. When the CONTAINER itself was stopped/rebooted/OOM-killed (which destroys the in-container tmux), the next launch RESUMES the last conversation from the bind-mounted transcript rather than starting fresh (durability model in Section 2, Key decision 1).
- I export a finished run's whole environment (toolchain plus workspace) to a portable, secret-free `.tar.gz`, move it to another machine, and import it back into a fresh case in one click.
- The container never accumulates: killing the session stops it, deleting the case removes it, and an instance-scoped boot reaper reaps containers whose case is gone.
Non-goals for the MVP: multi-tenant untrusted-code isolation guarantees (Codeman is loopback-default and single-operator, and the agent already runs `--dangerously-skip-permissions` on the host today), Kubernetes/compose orchestration, and per-command ephemeral containers.
## 2. Chosen architecture and why
The design grafts the strongest idea from each of the three proposals:
- Overlay-not-a-mode + faithful remote-SSH mirror (from "Docker Cases as a Location Overlay"): lowest churn, rides the existing quick-start / mux-sessions / state / recovery plumbing.
- Convenient-but-hardened default with an opt-in sealed profile, plus exec-time name-only secret env (from "Sealed Sandbox"): a strict security improvement over today's on-host execution without the UX tax of forcing an in-container re-login.
- One-artifact export + in-app import route (from "Container-as-Cargo"): the genuinely new, high-value capability Codeman lacks.
### Key decision 1: persistent per-CASE container, durable in-container tmux, AND resume-on-restart (the two-layer durability model)
Exactly one long-lived container per Docker case, named as a pure slug function `codeman-case-<slug>` (Docker charset `^[a-zA-Z0-9][a-zA-Z0-9_.-]+$`; Codeman already slugs case names for tmux), so create-if-missing and boot recovery are idempotent. PID1 is `sleep infinity` under `--init` (tini reaps zombies and forwards `docker stop`'s SIGTERM); the CLI is NOT the container command. The CLI runs inside a DURABLE in-container tmux on a dedicated socket `-L codeman-docker`, session `codeman-dkr-<id8>`, the direct analog of remote's `-L codeman-remote` / `codeman-ssh-<id8>`.
Two DIFFERENT failure surfaces need two DIFFERENT recovery layers, and conflating them is the central flaw the critic caught:
1. Codeman-PROCESS restart while the container stays up: the in-container tmux is still alive, so `tmux new-session -A` (attach-or-create) reattaches the SAME live agent and the paneCommand is ignored. This is the remote-SSH durability idiom and it works unchanged.
2. CONTAINER stop / daemon restart / host reboot / OOM-kill: the in-container tmux is GONE (fresh PID1). `new-session -A` will now CREATE a fresh session and run the paneCommand, which would start a brand-new conversation. This is the case the raw plan silently lost. Because the transcript directory is bind-mounted from the host (Key decision 3), the fix is to launch with RESUME: the paneCommand becomes `exec claude --dangerously-skip-permissions --resume <claudeSessionId>` (codex uses `resume <id>`, gemini `--resume <id>`) whenever a captured `claudeSessionId` exists. The `-A` semantics make this self-selecting: the resume flag only ever executes when tmux is actually re-created, which is exactly when the live session was lost. When tmux is still alive (case 1), attach wins and the flag is inert.
Capturing / persisting / reusing the resume id (the missing mechanism the critic flagged): Codeman already learns `Session.claudeSessionId` from transcript correlation (which works here because projHash matches, Key decision 3) and persists it in `SessionState`. We thread that value into `createSessionOptions` / `respawnPaneOptions` for docker so `buildDockerLaunchCommand` can inject the resume flag on any relaunch. To make a NEW Codeman session (new `id8`) re-launched against the same case resume its predecessor's conversation, we ALSO persist `lastClaudeSessionId` on the `DockerCase` record; the quick-start docker branch seeds the new `Session` with it when the `dockerResumeOnStart` setting is on. First-ever launch has no id, so it starts fresh. This is user-decision 7 (default resume behavior).
Reconciling with stop-on-kill and with the `--restart` policy (the internal inconsistency the critic found): the container is created with `--restart no` uniformly (Codeman's idempotent create-if-missing plus boot recovery is the single recovery mechanism; a restart policy would not preserve the conversation anyway because a restarted container gets a fresh PID1/tmux). Boot recovery re-runs `buildDockerLaunchCommand` from the restored `MuxSession.docker` (`docker inspect || docker create; docker start`, then exec with resume), so a host reboot or daemon restart recreates+starts the container and resumes the conversation instead of the session vanishing. `reconcileSessions` (tmux-manager.ts ~1800-1815) must NOT hard-delete a docker session merely because no LOCAL pane exists after the local `-L codeman` server died; docker (like remote) sessions are restored from `mux-sessions.json` and relaunched. This relaunch path is explicitly part of Phase 4/Phase 3 recovery work, not assumed.
Why this over the alternatives: `docker exec` gets SIGHUP and dies when its client TTY closes, so a bare `docker exec claude` restarts the CLI on every reconnect/respawn. The inner tmux plus resume is what makes reconnect idempotent across BOTH failure surfaces. Because this durability is the single most important design point, tmux-in-image is a HARD gated prerequisite (`checkDockerTmuxAvailable`), never a silent fallback to bare exec. Rejected alternatives: ephemeral-per-run or bare-exec containers (no reattach durability); a literal `'docker'` `SessionMode` (touches dozens of switch/enum sites and diverges from the remote overlay precedent, since Docker is a LOCATION orthogonal to the 5 CLI backends).
### Key decision 2: CLI + auth delivery
One prebuilt base image (built once, contains NO secrets): `node:22-bookworm-slim` + `git tmux ripgrep ca-certificates`, `npm i -g @anthropic-ai/claude-code @openai/codex @google/gemini-cli opencode-ai`, an `agent` user, HOME dirs made writable by an arbitrary host uid via the OpenShift "gid 0, group-writable" convention (Key decision 6). Because the toolchain is baked, export is reproducible and needs no network at import time. The image name/namespace/registry and its refresh cadence are user-decision 2 (the `codeman/agent:base` placeholder implies a Docker Hub org the project may not own).
Credentials are delivered ONLY at runtime, two commit-safe channels, default convenient:
- OAuth/config-file CLIs (Claude Max/Pro, gcloud, opencode): bind-mount the host credential dirs read-write (`~/.claude`, `~/.codex`, `~/.gemini` + `~/.config/gcloud`, `~/.config/opencode`) so the common user "just works" with no in-container login. Because these are bind mounts, `docker commit` (which captures only the container's own writable layer, never bind mounts) physically cannot capture them, so exports stay secret-free.
- API-key CLIs (codex/gemini): exec-time NAME-ONLY `docker exec --env OPENAI_API_KEY --env GEMINI_API_KEY ...` (no `=value`), sourced from Codeman's own process env. Only the key NAME appears in argv (no `ps` leak), and per-exec env is never captured by `docker commit`. This is the technique Codeman already uses via `tmux setenv` for the local Codex/Gemini panes, so it composes with existing machinery.
Per-host `DockerHost.mountCredentials` defaults `true` (convenient); setting it `false` yields a SEALED profile (no host cred mounts, in-container login only) for genuinely untrusted work. CRITICAL sealed-mode export rule (the leak the critic caught): in sealed mode the in-container login writes tokens into the container's OWN writable layer, which `docker commit` DOES capture, so a full-image export of a sealed container would ship credentials. Therefore full-image export is REFUSED for `mountCredentials:false` containers by default; the user may either take a workspace-only export (always safe) or opt into a pre-commit scrub that `docker exec`s `rm -rf ~/.claude ~/.codex ~/.gemini ~/.config/gcloud ~/.config/opencode` inside the container before commit (destructive to the in-container login, which is the point). This is enforced in the export route, not left to a manifest assertion.
Per-session `envOverrides` / `effort` / `codexConfig` / `geminiConfig` / `openCodeConfig` are REJECTED at quick-start exactly like the remote branch (session-routes.ts ~1698-1710). `modelOverride` is the one deliberate difference from remote: because the docker workspace is a REAL bind-mounted host dir that Codeman scaffolds (Key decision 5 and Section 6), `updateCaseModel()` can write the `model` key into `<workspace>/.claude/settings.local.json` and the in-container `claude` reads it, so the App Settings Claude Model picker works for docker cases. `effort` is a `--effort` CLI arg applied only by the local-spawn path we bypass, so it stays rejected (surfaced honestly in the UI, not silently inert). Per-mode command customization goes through `DockerHost.commands.<mode>` (`defaultDockerCommandForMode`, mirror of `defaultRemoteCommandForMode` at remote-hosts.ts:60). NEVER bake secrets into an image layer and NEVER pass a secret via create-time `-e` (both are committed).
Rejected alternative: sealed-by-default. For a single-operator loopback tool where the agent already runs skip-permissions on the host, forcing an in-container OAuth re-login is a UX regression with little real gain. We keep sealed as an opt-in. Rejected alternative: baking a login into the image, which leaks the instant you `docker save`.
### Key decision 3: workspace mount, container CWD, and transcript correlation
Bind-mount the host workspace dir into the container at the SAME absolute path (`dst == src`, mirror the host path), and set both `Session.workingDir` and the container workdir to that host path.
Two problems this solves that the raw proposals got wrong:
- File features: `DockerCase.hostWorkspacePath` is a REAL host directory, so `Session.workingDir = hostWorkspacePath` keeps file-routes, attachments, image-watcher, and previews working on real host bytes (unlike remote, where the path is remote-only and those features no-op). All three proposals wired `casePath = <container path>`; we deliberately diverge and use the host path.
- Transcript correlation: Claude writes transcripts under `~/.claude/projects/<hash-of-CWD>/`. By mirroring the host path as the container CWD, the projHash computed inside the container equals the host-side hash Codeman's transcript/subagent/workflow watchers expect, so correlation keeps working (and, in turn, feeds the resume-id capture in Key decision 1). A `/workspace`-style fixed dst would break it. Mirror-vs-fixed is user-decision 3.
`resolveMuxAttachCwd` still returns `/tmp` for docker sessions (the LOCAL bash pane only runs `docker exec`; it never needs the workspace as its cwd), mirroring remote.
### Key decision 4: network default and the engine-specific host gateway
Default `bridge` (own netns, NAT egress, no inbound), per-case changeable to `none` (offline shell sandbox; warned because it breaks the API CLIs) or `custom` (a user-defined bridge `codeman-net-<slug>`, the chokepoint for a future egress allowlist). `host` networking and any `-p` inbound publish are structurally unrepresentable in the flag builder and schema. Rationale: every API-backed CLI (Claude, Codex, Gemini) plus npm/git needs egress, so `bridge` is the only sane functional default; `none` is reserved for `shell`.
The host-callback gateway alias is ENGINE-SPECIFIC (the critic's podman finding): Docker uses `host.docker.internal`, Podman uses `host.containers.internal` (Docker's alias only exists on recent podman). A helper `hostGatewayAlias(engine)` returns the right name; Section 2.5, the create args, the `CODEMAN_API_URL` rewrite, and the host-guard allowlist all consume it, and BOTH aliases are added to the allowlist so a mixed fleet keeps working.
### Key decision 5: hooks actually reach the host AND are actually installed
Two independent things must both be true for a hook to fire, and the raw plan wired only the first:
1. Network reachability. Claude Code hooks POST to `$CODEMAN_API_URL` (`curl -sk`). Inside a bridge container `localhost` is the container and prod binds `127.0.0.1`, so we set `--add-host <gatewayAlias>:host-gateway` on create (skipped on Docker Desktop, where the alias is native), add the gateway alias to the host guard, and provide `CODEMAN_API_URL` and the hook secret (below).
2. Hook INSTALLATION. Hooks live in `<workspace>/.claude/settings.local.json`, written by the quick-start scaffolding block (around session-routes.ts ~1776) that calls `writeHooksConfig()` / `updateCaseModel()`. The raw plan extended the `!remote` guard to `!remote && !docker`, which would SKIP that block and silently disable ALL hooks regardless of networking. For docker the workspace is a REAL bind-mounted host dir, so the scaffolding block MUST run. Precise fix: extend to `!remote && !docker` ONLY the LOCAL-CLI-availability and local-spawn guards (the ones that stat the local binary or build the local spawn command); leave the workspace-scaffolding guard at `!remote` so it runs for docker. This same decision is what makes `modelOverride` work (Key decision 2). Consequence, surfaced as user-decision 4: linking a docker case now WRITES `.claude/settings.local.json` (and the CLAUDE.md scaffold, matching local-case behavior) into the user's real host directory, a behavioral shift from "link a dir" to "link and scaffold a dir."
`CODEMAN_API_URL` derivation (the wrong-scheme bug the critic caught): prod is HTTPS-only on 3000, and `server.ts` (~2000) auto-sets `process.env.CODEMAN_API_URL = ${protocol}://${apiHost}:${port}`. Hardcoding `http://host.docker.internal:3000` fails every hook. Instead a pure helper `containerApiUrl(process.env.CODEMAN_API_URL, engine)` parses the running URL and substitutes ONLY the hostname with `hostGatewayAlias(engine)`, preserving scheme and port (`https://host.docker.internal:3000`). Unit-tested against http, https, non-default ports, and both engines. Passed as create-time `--env CODEMAN_API_URL=<derived>` (case-stable, non-secret).
Hook secret and session attribution:
- `~/.codeman/hook-secret` is bind-mounted read-only to a container path; `--env CODEMAN_HOOK_SECRET_FILE=<that path>` is create-time (a path is non-secret; the bytes ride the bind mount and are never committed).
- `CODEMAN_SESSION_ID` (which the generated hooks reference at hooks-config.ts:78-80 to attribute events) plus `CODEMAN_MUX=1` are SESSION-scoped, so they are passed at EXEC time via `docker exec --env CODEMAN_SESSION_ID=<id> --env CODEMAN_MUX=1` (non-secret, value inline is fine, and exec env is not committed). Because a `tmux` session started fresh only inherits the invoking env when it starts the SERVER, the launch chain ALSO runs `tmux -L codeman-docker setenv -g CODEMAN_SESSION_ID <id>` (and `CODEMAN_MUX`) so reattaches and newly created panes see the same values. This mirrors how Codeman already injects per-session env into tmux for the external CLIs.
Hooks-in-MVP-vs-deferred stays user-decision 4; if deferred, docker ships as explicitly hook-degraded and we lean on output-based idle detection through the docker-exec PTY.
### Key decision 6: uid / HOME / rootless enforcement / macOS Docker Desktop
The raw plan showed `--user 1000:1000` in one place and `--user "$(id -u):$(id -g)"` in another and never resolved HOME writability; this section fixes all of it.
- Linux native (docker rootful or rootless): run `--user <hostUid>:0` (host uid, GID 0). The image follows the OpenShift arbitrary-uid convention: `HOME=/home/agent`, and `/home/agent` plus the tool cache dirs (`~/.npm`, `~/.cache`, `~/.config`) are owned `root:0` and group-writable (`chmod -R g+w`, `g+s` on dirs) so a process with GID 0 can write HOME even though its UID is not 1000. This keeps workspace files host-owned (the agent's UID is the host UID) AND keeps HOME writable, so the CLIs actually start.
- Podman rootless: use `--userns=keep-id` (maps the host uid to the image's `agent` uid inside the container) instead of `--user`, so `/home/agent` is owned by the running user and workspace files are host-owned. This is a real per-engine branch in `buildDockerCreateArgs`.
- macOS Docker Desktop: `--user <macUid>` (e.g. 501) does not own the image's `/home/agent`, so non-bind HOME writes fail EACCES and the CLIs may not start; Desktop also does its own bind-mount uid translation, provides `host.docker.internal` natively (no `--add-host`), and its VM memory ceiling can cap `--memory`. Detect Desktop via `docker info` (Server OS `linuxkit` / `OperatingString` contains "Docker Desktop") and take a dedicated path: do NOT pass `--user` (run as the image's baked `agent` uid and rely on Desktop's translation for workspace access), skip `--add-host`, and note in the UI that memory caps are subject to the VM ceiling.
Rootless resource-cap enforcement (the silently-inert risk): rootless Docker without cgroup-v2 systemd delegation (`Delegate=yes`) silently IGNORES `--memory`/`--cpus`/`--pids-limit`. The probe checks `docker info` for `CgroupVersion=2` plus rootless plus delegation; if caps cannot be enforced, `checkDockerAvailable` returns `capsEnforced:false` and the link/probe surfaces "resource caps are advisory on this engine." Whether to REQUIRE delegation or ship-with-warning is user-decision 6.
## 3. Data model
New TypeScript types in `src/types/session.ts`, added right after the remote types (lines 46-99). SessionMode (line 44) is UNCHANGED.
```ts
export type DockerCommandMode = Extract<SessionMode, 'shell' | 'claude' | 'opencode' | 'codex' | 'gemini'>;
export type DockerEngine = 'docker' | 'podman';
export type DockerNetworkMode = 'bridge' | 'none' | 'custom'; // never 'host'
export interface DockerResourceLimits {
memory?: string; // '4g' -> --memory 4g --memory-swap 4g (swap==memory: real OOM cap)
cpus?: string; // '2'
pidsLimit?: number; // 512 (fork-bomb guard)
nofile?: string; // '4096:8192'
shmSize?: string; // optional; only when a tool needs /dev/shm
}
export interface DockerHost {
id: string;
label: string;
engine?: DockerEngine; // default resolved by probe (docker, else podman)
image: string; // default resolved image ref (see user-decision 2)
daemonHost?: string; // advanced: -H ssh://user@host / DOCKER_HOST
context?: string; // advanced: --context <ctx>
network?: DockerNetworkMode; // default 'bridge'
networkName?: string; // when network === 'custom'
resources?: DockerResourceLimits;
mountCredentials?: boolean; // default true (false = sealed; blocks full-image export)
hooksEnabled?: boolean; // default true (host-gateway callback wiring)
resumeOnStart?: boolean; // default true (see Key decision 1 / user-decision 7)
commands?: Partial<Record<DockerCommandMode, string>>;
extraCreateArgs?: string[]; // validated like extraSshOptions
extraExecArgs?: string[];
}
export interface DockerCase {
name: string;
type: 'docker';
hostId: string;
hostWorkspacePath: string; // absolute HOST dir: bind src + Session.workingDir
containerWorkdir?: string; // container path; default = hostWorkspacePath (mirror -> projHash match)
container?: string; // default codeman-case-<slug>
lastClaudeSessionId?: string; // captured resume id (Key decision 1)
}
export interface SessionDocker { // flattened, round-trips through mux/state (mirror SessionRemote at 91)
hostId: string;
label: string;
engine: DockerEngine;
image: string;
containerName: string;
hostWorkspacePath: string;
containerWorkdir: string;
network: DockerNetworkMode;
networkName?: string;
resources?: DockerResourceLimits;
mountCredentials: boolean;
hooksEnabled: boolean;
resumeOnStart: boolean;
daemonHost?: string;
context?: string;
commands?: Partial<Record<DockerCommandMode, string>>;
extraCreateArgs?: string[];
extraExecArgs?: string[];
configHash?: string; // drift detection (Key decision, Section 4)
}
```
- `SessionState` gains `docker?: SessionDocker` immediately after `remote?` (line 219). It persists automatically because `SessionState` is structural and `state-store.ts` stores `toState()` verbatim.
- `src/mux-interface.ts`: add `docker?: SessionDocker` to `MuxSession` (after line 38), `CreateSessionOptions` (after 81), `RespawnPaneOptions` (after 105). `MuxSession.docker` round-trips through `mux-sessions.json` automatically.
- `src/types/api.ts` `CaseInfo`: add `'docker'` to the `location` union and a `docker?: { hostId; container; image?; path; network }` display block.
- `src/services/unified-session-service.ts`: add a boolean `docker?` flag on `UnifiedSessionItem` and source rows, set from `MuxSession.docker` presence (mirror the `remote` flag at ~line 200 and the harvest at session-routes.ts:2313).
New state files (all via `dataPath()`, mirroring `remote-hosts.json` / `remote-cases.json`):
- `~/.codeman/docker-hosts.json` (reusable engine/image/network/resource profiles).
- `~/.codeman/docker-cases.json` (`name -> DockerCase`, including `lastClaudeSessionId`).
- `~/.codeman/docker-exports/` (dedicated dir for `.image.tar.gz` + `.workspace.tar.gz` + `manifest.json`; never inline in state.json; retention/pruning per Section 5).
No new `state.json` / `mux-sessions.json` files: `SessionState.docker` and `MuxSession.docker` ride the existing serialization.
## 4. Container lifecycle (exact command shapes)
All builders are PURE string functions (directly unit-testable). Host values interpolated into the outer `bash -c "..."` layer (container name, image, workdir, host paths) are `shellescape()`'d and, for user-supplied fields, schema-rejected for `$`/backtick via `NO_SHELL_META`. The escaping chain here is DEEPER than remote's single `ssh '<tmux ...>'`: the whole `docker inspect || docker create <dozens of --mount/--env/shellescaped host paths>` is interpolated into `bash -c "..."` then `JSON.stringify`'d into respawn-pane. This is a known place to get stuck, so it is covered by concrete escaping tests (Section 9), including host workspace paths containing spaces, not just a "we call shellescape" claim.
New in `src/tmux-manager.ts`:
```ts
const DOCKER_TMUX_SOCKET = 'codeman-docker';
// 'dkr' letters deliberately FAIL SAFE_MUX_NAME_PATTERN (^codeman-[a-f0-9-]+$),
// so a Codeman running INSIDE the container never adopts/resizes/respawns our session.
export function dockerTmuxSessionName(id: string): string { return `codeman-dkr-${id.slice(0, 8)}`; }
```
`buildDockerBaseArgs(docker)` (pure, in `docker-hosts.ts`, mirror of `buildSshConnectionArgs`) emits the engine prefix tokens: `docker` (or `podman`) + optional `--context <ctx>` or `-H <daemonHost>`. `buildDockerCreateArgs(docker, sessionId)` emits the `docker create` flag array (with the per-engine uid/userns branch from Key decision 6).
IMAGE PRESENCE (before any create, the auto-pull footgun the critic caught): the launch chain runs `docker image inspect <image> >/dev/null 2>&1` first; on miss it exits with a distinct message ("base image <ref> not present: build with scripts/build-agent-image.mjs or pull it") rather than triggering a blocking multi-GB auto-pull inside the tmux pane. `docker create` carries `--pull=never`. The tmux-availability probe likewise uses `docker run --rm --pull=never <image> sh -lc 'command -v tmux'` and reports the same build/pull hint if the image is absent, so the 15s-bounded probe never hangs on a pull.
CREATE (the ensure step, embedded in the launch string):
```
docker create \
--name codeman-case-myproj --hostname myproj \
--label codeman.managed=1 --label codeman.instance=<CODEMAN_INSTANCE> \
--label codeman.case=myproj --label codeman.session=<id8> \
--label codeman.confighash=<hash> \
--pull=never --init --restart no \
--user 1000:0 \
--workdir '/home/arkon/cases/myproj' \
--mount type=bind,src='/home/arkon/cases/myproj',dst='/home/arkon/cases/myproj' \
--mount type=bind,src='/home/arkon/.claude',dst='/home/agent/.claude' \
--mount type=bind,src='/home/arkon/.codeman/hook-secret',dst='/home/agent/.codeman/hook-secret',readonly \
--add-host host.docker.internal:host-gateway \
--memory 4g --memory-swap 4g --cpus 2 --pids-limit 512 --ulimit nofile=4096:8192 \
--cap-drop ALL --security-opt no-new-privileges \
--network bridge \
--env HOME=/home/agent --env TERM=xterm-256color --env COLORTERM=truecolor \
--env CODEMAN_API_URL=https://host.docker.internal:3000 \
--env CODEMAN_HOOK_SECRET_FILE=/home/agent/.codeman/hook-secret \
codeman/agent:base \
sleep infinity
```
- `--user 1000:0` shown is the Linux-native form with GID 0 (Key decision 6); it is actually `--user <hostUid>:0`, or `--userns=keep-id` for podman rootless, or omitted on Docker Desktop. The literal is illustrative only.
- Create-time `--env` carries only NON-SESSION, non-secret, case-stable values (safe to be committed): the DERIVED `CODEMAN_API_URL` (https-preserving, Key decision 5) and the hook-secret FILE PATH. `CODEMAN_SESSION_ID`/`CODEMAN_MUX` and the codex/gemini key NAMES are exec-time only.
- `codeman.instance=<CODEMAN_INSTANCE>` is REQUIRED on the label set so the boot reaper is instance-scoped (a beta/second instance must never reap prod's containers).
- `codeman.confighash` is a stable hash of the drift-relevant create args (image, resources, network, mounts, non-session env). Drift detection (user story 2, the config-never-takes-effect gap): on launch the ensure block compares the desired hash to the existing container's label; on mismatch the launch does NOT silently reuse the stale container. Instead the docker route returns a "container config changed, recreate?" action (SSE + UI confirm), and on confirm Codeman `docker rm`'s and recreates. rm destroys in-image (non-bind) state, but the workspace and transcripts survive on their bind mounts and the conversation is restored via `--resume`, so the recreate is safe. Auto-recreate-vs-prompt is a UI choice; the MVP prompts.
- `--restart no` (resolved consistently with Key decision 1; recovery is Codeman's idempotent create-if-missing, not an engine restart policy, which also matters for Podman which has no daemon).
EXEC (`buildDockerLaunchCommand`, the docker analog of `buildRemoteLaunchCommand`, TTY-correct, resume-aware). The whole thing is ONE `bash -c` string that image-checks, ensures, starts, primes tmux env, then execs:
```
docker image inspect codeman/agent:base >/dev/null 2>&1 || { echo 'Codeman: base image codeman/agent:base not present (build or pull it)'; exit 1; } ; \
docker inspect codeman-case-myproj >/dev/null 2>&1 || docker create <all create args above> ; \
docker start codeman-case-myproj >/dev/null 2>&1 || { echo 'Codeman: container codeman-case-myproj failed to start (daemon down?)'; exit 1; } ; \
exec docker exec -it \
--workdir '/home/arkon/cases/myproj' \
--env TERM=xterm-256color --env COLORTERM=truecolor \
--env CODEMAN_SESSION_ID=1a2b3c4d --env CODEMAN_MUX=1 \
--env OPENAI_API_KEY --env GEMINI_API_KEY \
codeman-case-myproj \
sh -lc 'tmux -L codeman-docker setenv -g CODEMAN_SESSION_ID 1a2b3c4d \; setenv -g CODEMAN_MUX 1 \; new-session -A -s codeman-dkr-1a2b3c4d -c '\''/home/arkon/cases/myproj'\'' '\''cd /home/arkon/cases/myproj && exec claude --dangerously-skip-permissions --resume <claudeSessionId>'\'' \; set -t codeman-dkr-1a2b3c4d status off \; set -t codeman-dkr-1a2b3c4d mouse off \; set -t codeman-dkr-1a2b3c4d prefix C-q \; set -s escape-time 0'
```
- `docker exec -it`: `-t` allocates a PTY and forwards SIGWINCH into the container so the Ink TUI re-lays-out on pane resize; `TERM`/`COLORTERM` prevent degraded rendering. `--env OPENAI_API_KEY` (name only) is present only for codex/gemini and is exec-time (never committed). `CODEMAN_SESSION_ID`/`CODEMAN_MUX` are exec-time values plus a `tmux setenv -g` prime so reattaches and new panes inherit them (Key decision 5).
- `--resume <claudeSessionId>` (codex `resume <id>`, gemini `--resume <id>`) is appended to `modeCommand` ONLY when a captured id exists; on first launch it is omitted. `new-session -A` makes the flag inert on a live-tmux reattach and effective only when tmux is re-created (Key decision 1).
- `modeCommand = docker.commands?.[mode] || defaultDockerCommandForMode(mode)` (`exec claude --dangerously-skip-permissions`, `exec bash -l`, etc.), with the resume suffix injected by the builder.
- Escaping survives every layer identically to remote in shape but deeper in nesting: `paneCommand` (`cd ... && exec ...`) is one shellescaped tmux arg, the whole `tmuxInvocation` is one shellescaped `sh -lc` arg, and the outer string is `JSON.stringify()`'d into `bash -c` by respawn-pane (tmux-manager.ts:1329).
Wire-up (extend the two existing seams to 3-way):
- createSession (tmux-manager.ts:1276): `const fullCmd = docker ? buildDockerLaunchCommand({ mode, docker, sessionId, resumeSessionId }) : remote ? buildRemoteLaunchCommand({ mode, remote, sessionId }) : localFullCmd;`
- launchCmd cd-skip (tmux-manager.ts:1327): `const launchCmd = (remote || docker) ? fullCmd : \`cd ${JSON.stringify(workingDir)} && ${fullCmd}\`;`
- respawnPane: same two edits at lines 1524 and 1542.
START / reattach-after-reboot: the ensure block (image-check, `docker inspect || docker create`, `docker start`) is fully idempotent, so boot recovery just re-runs `buildDockerLaunchCommand` from the restored `MuxSession.docker` with the persisted resume id. A rebooted host recreates the container and resumes the conversation.
DOCKER-DOWN surfacing (the PTY-exit-breaker false-trip risk): if `docker start` or `docker exec` cannot attach (daemon down, container missing), the launch prints a docker-specific message and exits, which alone would still count toward `session-pty-exit-breaker` and show a generic "respawn breaker tripped" push. To avoid masking the cause, the docker reattach path runs a fast `checkDockerAvailable` pre-flight: if the daemon/container is unreachable, Codeman broadcasts a docker-specific error (SSE + push, "container <name> is not running / daemon down") and SKIPS the auto-reattach that would trip the breaker, rather than fast-looping `docker exec`.
STOP / KILL (`killSession` Strategy 3c, right after remote's Strategy 3b at tmux-manager.ts:1719, guarded by `IS_TEST_MODE`):
```ts
if (session.docker) {
// best-effort, fire-and-forget, timeout-bounded so it never blocks the local kill
execAsync(buildDockerKillCommand({ docker: session.docker, sessionId }), { timeout: EXEC_TIMEOUT_MS }).catch(() => {});
}
```
`buildDockerKillCommand` emits: `docker exec codeman-case-<slug> tmux -L codeman-docker kill-session -t codeman-dkr-<id8> ; docker stop -t 10 codeman-case-<slug>`. Stopping frees CPU/RAM and, per Key decision 1, is safe for conversation continuity because the NEXT launch resumes from the bind-mounted transcript via `--resume`. Whether to stop at all (RAM vs instant live-agent reattach) is user-decision 6/1 (reframed honestly). The bind-mounted workspace and transcripts always survive on the host.
REMOVE: only on explicit case delete (`docker rm -f codeman-case-<slug>`), gated behind an "export first?" UI prompt because rm destroys any in-image (non-bind) state. Instance-scoped boot reaper (fixing the racy/cross-instance reaper): after `docker-cases.json` is loaded AND after `restoreMuxSessions` has run, enumerate `docker ps -a --filter label=codeman.managed=1 --filter label=codeman.instance=<CODEMAN_INSTANCE> --format '{{.Names}}\t{{index .Labels "codeman.case"}}'` and `docker rm -f` only containers whose case is gone from THIS instance's `docker-cases.json`. The instance filter is what stops a beta reaping prod's containers (the exact cross-instance hazard the project memory warns about).
AVAILABILITY PROBE (`docker-hosts.ts`, timeout-bounded like `checkRemoteTmuxAvailable`'s 15s, `IS_TEST_MODE` no-op):
```
docker info --format '{{json .}}' # server up, CgroupVersion, rootless, OS (Desktop detect), cap-delegation
docker image inspect <image> --format '{{.Id}}' # image PRESENT (no auto-pull)
docker run --rm --pull=never <image> sh -lc 'command -v tmux' # tmux-in-image gate (hard prerequisite), only if image present
```
`checkDockerAvailable()` returns `{ ok, engine, rootless, isDesktop, cgroupV2, capsEnforced }` (parse `SecurityOptions` for `name=rootless`, `CgroupVersion`, delegation, and Server OS for Desktop). `checkDockerTmuxAvailable(host)` returns a structured result with a user-facing error and correct install hint (NOT `npm install -g`; the hint is "build/pull the base image" for a missing image and "install docker or podman" for a missing engine).
IN-CONTAINER CLI VERSION (fixing the #154 wheel-forwarding regression): the raw plan skipped the LOCAL `cliVersion` probe for docker (correct, since it reports the HOST claude) but left `cliVersion` undefined, which disables trackpad wheel-forwarding. Instead, for docker sessions Codeman runs an IN-CONTAINER probe `docker exec <container> claude --version` (bounded, `IS_TEST_MODE` no-op) and feeds THAT into `cliVersion`. This also means a stale baked CLI is visible; combined with the rebuild-cadence in user-decision 2, agents are not silently pinned to an old claude.
## 5. Export / Import
EXPORT is a concurrency-bounded job (reuse `runWithConversionLimit` from `document-conversion-limiter.ts` so N simultaneous exports cannot fork-bomb the host). Route `POST /api/docker-cases/:name/export`.
Preconditions (the consistency and leak risks the critic caught):
- Sealed guard: if `mountCredentials:false`, full-image export is REFUSED unless the caller explicitly opts into the pre-commit scrub (Key decision 2). Workspace-only export is always allowed.
- Quiesce + free-space: require the session idle, then `docker pause` the container spanning BOTH the workspace tar AND the commit so the two artifacts are mutually consistent (the raw plan paused only the commit, leaving the bind-mount tar to run against a mid-write agent). Before any heavy step, precheck free space in the exports dir and in `/var/lib/docker`; if below `DOCKER_EXPORT_MIN_FREE_BYTES`, refuse with a clear error (a full `/var/lib/docker` wedges the daemon and breaks EVERY session on the host).
Steps (all cleanup in try/finally so a mid-way failure never orphans an intermediate image or leaves the container paused):
1. `docker commit -c 'LABEL codeman.exported=1' codeman-case-<slug> codeman/export-<slug>:<ts>` (unique tag per export defeats the stale-image trap). Optional pre-commit scrub in sealed mode as above; also blank instance-specific committed env (`-c 'ENV CODEMAN_API_URL='` etc.) so the image carries no stale host references.
2. `docker save codeman/export-<slug>:<ts> | gzip` streamed in fixed 8192-byte chunks to `~/.codeman/docker-exports/<slug>-<ts>.image.tar.gz`. Uses `docker save` (layers + repo:tag + CMD), never `docker export` (flat rootfs), so restore is a trivial `docker load`.
3. `tar --numeric-owner -C <hostWorkspacePath> -czf <slug>-<ts>.workspace.tar.gz .` while paused (the bind-mounted workspace is NOT in the image, so it travels separately and consistently).
4. Write `manifest.json`: schema version, caseName, image tag, engine, containerWorkdir, resource/network config, codeman version, base-image digest, createdAt, per-member sha256, `mountCredentials`, and `secretFree` (true only for convenient-mode or scrubbed-sealed exports).
5. `docker rmi codeman/export-<slug>:<ts>` in the `finally` (delete the intermediate committed image regardless of success), then `docker unpause`.
The three files are wrapped in one bundle `<slug>-<ts>.codeman-container.tgz` and offered as a downloadable artifact through the existing file-routes streaming + attachment-registry handoff.
Retention / disk budget (user-decision 3): `docker-exports/` is capped at `DOCKER_EXPORT_KEEP` most-recent bundles with an auto-prune on each new export, plus the free-space precheck above. Workspace scrub: the WORKSPACE tar gets a scan/warn pass for agent-created `.env` / `.git/credentials` (a distinct leak channel from container creds). A lighter "workspace-only" export (just the workspace tar, no commit/save) is the fast default for 24h+ runs; full-image is the explicit heavier option (user-decision 7 in the original list, now decision on the default button below).
What travels: the baked toolchain image plus any in-image writes, and the workspace tar. What does NOT travel: bind-mounted credentials (physically excluded from commit) and anything that lived only in a bind mount. Secret-free by construction in convenient mode, and enforced (refuse-or-scrub) in sealed mode.
IMPORT `POST /api/docker-cases/import` (untrusted-bundle containment, the traversal/overwrite risk): stream the uploaded bundle, validate every manifest checksum BEFORE any extraction or load. Extract the workspace tar with `tar --no-absolute-names -C <fresh dir>` PLUS per-entry validation rejecting any member whose normalized path escapes the destination (leading `/` or `..` components). `gunzip | docker load` the image, then RE-TAG the loaded image id into a quarantined namespace `codeman/imported-<slug>:<ts>` and NEVER allow the load to overwrite `codeman/agent:base` or any pre-existing tag (capture the loaded id, ignore the bundle's repo:tag). Create a NEW `DockerCase` pointing at the quarantined image with THIS host's mounts/creds and the manifest's resource/network config, and recreate the container hardened (cap-drop ALL, no-new-privileges, non-root, `--pull=never`, CMD overridden to `sleep infinity`). The destination supplies its own login, so credentials never cross machines. Plus `GET /api/docker-exports` (list) and `DELETE /api/docker-exports/:filename`, all behind Codeman's existing auth / loopback-default / host-guard / Origin-CSRF stack.
## 6. Codeman integration (file-by-file, mirroring the remote-SSH feature)
- `src/types/session.ts`: add `DockerCommandMode`, `DockerEngine`, `DockerNetworkMode`, `DockerResourceLimits`, `DockerHost`, `DockerCase`, `SessionDocker` (Section 3). Add `docker?: SessionDocker` to `SessionState` after line 219. SessionMode (line 44) UNCHANGED.
- `src/mux-interface.ts`: add `docker?: SessionDocker` to `MuxSession` (38), `CreateSessionOptions` (81), `RespawnPaneOptions` (105).
- `src/docker-hosts.ts` (NEW, direct mirror of `src/remote-hosts.ts`): `readDockerHosts`/`writeDockerHosts`/`readDockerCases`/`writeDockerCases` (via `dataPath`, including `lastClaudeSessionId` read/write), `defaultDockerCommandForMode` (mirror line 60), `dockerDisplayPath` (`container:/path`, mirror `remoteDisplayPath` at 205), `toSessionDocker(host, case)` (mirror `toSessionRemote` at 212), `buildDockerBaseArgs`/`buildDockerCreateArgs` (per-engine uid/userns branch), `hostGatewayAlias(engine)`, `containerApiUrl(processApiUrl, engine)` (scheme+port-preserving, unit-tested), `checkDockerAvailable`/`checkDockerTmuxAvailable`/`probeDockerCliVersion` (15s-bounded, `IS_TEST_MODE` no-op), a config-hash helper for drift, its own POSIX `shellescape` copy (mirror line 83). `const IS_TEST_MODE = !!process.env.VITEST;` gates every real `docker` invocation.
- `src/tmux-manager.ts`: add `DOCKER_TMUX_SOCKET`, `dockerTmuxSessionName`, `buildDockerLaunchCommand` (resume-aware, image-check, env-prime), `buildDockerKillCommand` (Section 4). Extend the two `fullCmd` ternaries (1276, 1524) and the two `launchCmd` cd-skips (1327, 1542). Add `killSession` Strategy 3c after 1719. Ensure `reconcileSessions` (~1800-1815) does NOT hard-delete docker sessions on local-tmux death (recovery relaunch path).
- `src/session.ts`: add `_docker?: SessionDocker` field (mirror `_remote` at 403), constructor arg (477), assignment (550). Thread `docker: this._docker` and `resumeSessionId: this._claudeSessionId` into BOTH `createSessionOptions` and `respawnPaneOptions` in `startInteractive` (1352/1370) and the second path (1740/1750). Emit `docker: this._docker` in `toState()` (1010). Replace the LOCAL cliVersion probe at 1320 for docker with the IN-CONTAINER `probeDockerCliVersion` (do not merely skip it). Extend `resolveMuxAttachCwd(workingDir, remote, docker)` (215) to return `/tmp` when `docker` is set. On claudeSessionId capture, persist it to the owning `DockerCase.lastClaudeSessionId`.
- `src/web/server.ts`: in `restoreMuxSessions` (2160), add `docker: muxSession.docker ?? savedState?.docker` to the `new Session({...})` call (2195-2216), and skip docker in the same `isExternalCliMode`/Ralph recovery guards as remote. Register the instance-scoped boot reaper to run AFTER docker-cases load and AFTER `restoreMuxSessions`. Ensure `CODEMAN_API_URL` derivation reads the SAME `process.env.CODEMAN_API_URL` the server sets at ~2000.
- `src/web/schemas.ts`: add `DockerHostSchema` and `DockerCaseLinkSchema` (below). The three mode enums (177/373/705) and `QuickStartSchema` (368) UNCHANGED (docker resolves by `caseName` lookup like remote).
- `src/web/routes/session-routes.ts`: import the docker helpers from `../../docker-hosts.js`. Add a docker branch in `/api/quick-start` parallel to the remote branch (1686-1720): `readDockerCases` -> find by `caseName` -> `readDockerHosts` -> find by `hostId`; reject `envOverrides`/`effort`/`codexConfig`/`geminiConfig`/`openCodeConfig` (but ACCEPT `modelOverride`, which flows via scaffolded `settings.local.json`); run `checkDockerAvailable` + `checkDockerTmuxAvailable` (image-present, engine, caps-enforced); surface `capsEnforced:false` and Desktop notes; set `casePath = dockerCase.hostWorkspacePath` (REAL host dir), `docker = toSessionDocker(host, dockerCase)`, and seed `resumeSessionId` from `dockerCase.lastClaudeSessionId` when `resumeOnStart`. Extend the LOCAL-availability and local-spawn guards (around 1796/1810) to `!remote && !docker`, but DO NOT extend the workspace-scaffolding guard (~1776, `writeHooksConfig`/`updateCaseModel`), which MUST run for docker. Pass `docker` into `new Session` (1847); `autoConfigureRalph` (1853) gated on `!docker`. Add `docker: m.docker !== undefined ? true : undefined` to the unified harvest (2313).
- `src/web/routes/case-routes.ts`: import the docker read/write/check helpers + schemas. Add a docker listing loop in `GET /api/cases` (mirror 94-119, `location: 'docker'`, `docker: {...}` via `dockerDisplayPath`). Add `/api/docker-hosts` GET/POST/PUT/DELETE (mirror 168-204) and `POST /api/cases/docker-link` (mirror 206-232; run `checkDockerAvailable`/`checkDockerTmuxAvailable` at link time; broadcast `CaseLinked` with `type: 'docker'`). Add a docker-unlink branch to `DELETE /api/cases/:name` (mirror 288-296; `docker rm -f`; broadcast `CaseDeleted` `type: 'docker-unlinked'`). Add the docker branch to single-case `GET` (mirror 358-368). Add `POST /api/docker-cases/:name/export`, `/import`, `GET/DELETE /api/docker-exports`, and a `POST /api/docker-cases/:name/recreate` (drift confirm) per Sections 4 and 5.
- `src/web/sse-events.ts` + `src/web/public/constants.js`: reuse `CaseLinked`/`CaseDeleted` for CRUD. Add `docker:exportProgress`, `docker:exportComplete`, `docker:importComplete`, `docker:configDrift`, and `docker:containerError` to BOTH registries (kept in sync per CLAUDE.md).
- Frontend `src/web/public/index.html` (~1831): add a Docker `modal-tab-btn` next to Remote; add a `#case-docker` panel mirroring `#case-remote` with `dockerCaseName`, `dockerHostWorkspacePath`, `dockerContainer`, `dockerImage`, `dockerHostId`, and an Advanced `<details>` for network mode, resource caps, `mountCredentials`, `resumeOnStart`, and remote daemon. Surface a "scaffolds .claude into this host dir" note (user-decision 4) and a "resource caps advisory on this engine" warning when `capsEnforced:false`.
- Frontend `src/web/public/session-ui.js`: `formatCasePickerLabel` (48) + `buildCasePickerOptions` (71-73) handle `location === 'docker'` (`name @ container`, add container/image to the search haystack); `resetCaseModalFields` (~1514) add a `dockerFields` array; `switchCaseModalTab` (1573/1580/1597) handle `'case-docker'`; `submitCaseModal` add the docker branch; new `linkDockerCase()` (mirror `linkRemoteCase` at 1689) POSTing `/api/docker-hosts` then `/api/cases/docker-link`, sending omitted optionals as `undefined` (spread `...(x ? {x} : {})`, never `null`, per the Zod `.optional()`-rejects-null gotcha); `runClaude` (520) / `runShell` (702) extend the `location === 'remote'` routing to also match `'docker'`; `runOpenCode`/`runCodex`/`runGemini` (792/846/900) make the `isRemote` checks `isRemoteOrDocker` so local status probes are skipped. In the session-options Summary tab, note that `effort` is inert for docker (rejected) while `model` IS honored via `settings.local.json`.
- Frontend `src/web/public/panels-ui.js` (425-426): add `caseItem?.docker?.path`/`container` to the case-search fields.
Schemas (`src/web/schemas.ts`), mirroring `RemoteHostSchema` (299) / `RemoteCaseLinkSchema` (351):
```ts
export const DockerHostSchema = z.object({
id: z.string().regex(/^[a-zA-Z0-9_-]+$/, 'Invalid docker host id'),
label: z.string().min(1).max(100),
engine: z.enum(['docker', 'podman']).optional(),
image: z.string().min(1).max(512).regex(/^[a-zA-Z0-9][\w./:@-]*$/, 'Invalid image ref').regex(NO_SHELL_META),
daemonHost: z.string().max(512).regex(NO_SHELL_META, 'Invalid daemon host').optional(),
context: z.string().max(128).regex(/^[a-zA-Z0-9._-]+$/, 'Invalid context').optional(),
network: z.enum(['bridge', 'none', 'custom']).optional(),
networkName: z.string().max(128).regex(/^[a-zA-Z0-9][a-zA-Z0-9_.-]+$/).optional(),
resources: z.object({
memory: z.string().regex(/^\d+[bkmg]?$/i).optional(),
cpus: z.string().regex(/^\d+(\.\d+)?$/).optional(),
pidsLimit: z.number().int().positive().max(100000).optional(),
nofile: z.string().regex(/^\d+:\d+$/).optional(),
shmSize: z.string().regex(/^\d+[bkmg]?$/i).optional(),
}).strict().optional(),
mountCredentials: z.boolean().optional(),
hooksEnabled: z.boolean().optional(),
resumeOnStart: z.boolean().optional(),
commands: RemoteCommandOverridesSchema, // reuse the shared shape
extraCreateArgs: z.array(z.string().min(1).max(1024).regex(NO_SHELL_INJECTION).refine(noCommandSubstitution)).max(32).optional(),
extraExecArgs: z.array(z.string().min(1).max(1024).regex(NO_SHELL_INJECTION).refine(noCommandSubstitution)).max(32).optional(),
});
export const DockerCaseLinkSchema = z.object({
name: z.string().regex(/^[a-zA-Z0-9_-]+$/, 'Invalid case name format'),
hostId: z.string().regex(/^[a-zA-Z0-9_-]+$/, 'Invalid docker host id'),
hostWorkspacePath: z.string().min(1).max(2000).regex(/^\//, 'Path must be absolute').regex(NO_SHELL_META, 'Invalid characters in workspace path'),
containerWorkdir: z.string().min(1).max(2000).regex(/^\//).regex(NO_SHELL_META).optional(),
container: z.string().min(2).max(128).regex(/^[a-zA-Z0-9][a-zA-Z0-9_.-]+$/, 'Invalid container name').optional(),
});
```
`NO_SHELL_META` (rejects `$`/backtick, schemas.ts:297) is REQUIRED on `image`, `hostWorkspacePath`, `containerWorkdir`, and `container`, because all four reach the outer `bash -c "..."` double-quote layer where `$(...)`/backtick re-expose, exactly the reason `remotePath`/`identityFile` use it. `--privileged` and any `-v /var/run/docker.sock` are structurally unrepresentable (never emitted by the builder, never accepted by the schema).
## 7. Security model
- Hardening flags on every create: `--cap-drop ALL`, `--security-opt no-new-privileges` (NOT auto-set by rootless Docker or Podman, so always explicit), the uid/userns branch of Key decision 6 (never container-root; workspace files stay host-owned and HOME stays writable via GID 0), `--pids-limit` (fork-bomb guard), `--memory` with `--memory-swap == --memory` (real OOM cap), `--ulimit nofile`, `--init`, `--pull=never`. NEVER `--privileged`, NEVER mount the docker socket into the agent container. `--storage-opt size=` is emitted ONLY after the probe confirms overlay2-on-xfs-pquota or btrfs (the AICE-class silently-ignored trap); otherwise it is omitted and the UI does not advertise a size cap. Resource caps are advertised as ENFORCED only when the probe reports `capsEnforced:true`; under non-delegated rootless they are labeled advisory (user-decision 6).
- Engine: prefer whichever the probe finds, Podman-rootless first for security (a container-root breakout lands as an unprivileged host user). Rootless bind-mount ownership uses `--userns=keep-id` (Podman) vs `--user <hostUid>:0` (Docker), so real per-engine branching lives in `buildDockerCreateArgs`. Docker Desktop takes its own uid path (Key decision 6).
- Blast radius (the combined-posture the critic asked to surface, user-decision 5): the default convenient profile mounts an arbitrary host workspace dir RW (host-owned, mirrored path) AND host `~/.claude`/`~/.codex`/`~/.gemini`/`~/.config/gcloud`/`~/.config/opencode` RW into a NETWORK-ENABLED container. Container-run agent code can therefore read/modify those host trees and reach the network simultaneously. This is still a strict improvement over today's on-host skip-permissions execution, but the user must accept the combined posture explicitly; the sealed profile plus `network:none` is the mitigation for genuinely untrusted work.
- Secret handling: creds arrive ONLY as bind-mounted files (default) or exec-time NAME-ONLY `--env` (codex/gemini keys), NEVER as create-time `-e` and NEVER as an image layer. Sealed-mode export is refuse-or-scrub (Section 5), closing the sealed-leak inversion.
- CLAUDE.md "Multi-CLI prefix discipline": the exec-time name-only env is restricted to the CLI-specific keys per mode (Claude: none with OAuth mount; Codex: `OPENAI_API_KEY`/`CODEX_API_KEY`; Gemini: `GEMINI_API_KEY`/`GOOGLE_*`), never a blanket forward. `envOverrides` is rejected for docker, so the `ALLOWED_ENV_PREFIXES` allowlist is not widened.
- hook-secret: bind-mounted read-only, referenced via `CODEMAN_HOOK_SECRET_FILE` (a path, non-secret); the secret bytes never enter env or the image. Both `host.docker.internal` and `host.containers.internal` are added to the host-guard allowlist so the in-container hook curl's Host header passes on either engine.
- Host guard / instance isolation: the in-container tmux socket (`codeman-docker`) and name (`codeman-dkr-<id8>`) deliberately FAIL a container-internal Codeman's `SAFE_MUX_NAME_PATTERN`, so a nested Codeman never adopts our session (unit-asserted). The boot reaper is instance-scoped by the `codeman.instance` label so a beta never reaps prod. Any remote-daemon (`-H`/`--context`) mode is host-root-equivalent and stays strictly behind the existing auth/loopback/host-guard/Origin-CSRF stack.
- Import containment: untrusted bundles are checksum-validated, extracted with traversal guards, and loaded into a quarantined image namespace (never overwriting the base image), then run with the same hardening.
## 8. Phased implementation (branch: `feat/docker-session-mode`)
Each phase is independently testable; per CLAUDE.md, end-to-end test in the real env before COM. All new docker IO paths carry `const IS_TEST_MODE = !!process.env.VITEST;` and no-op under it; the pure command builders are tested directly.
- Phase 0: base image + engine probe. Author `docker/agent.Dockerfile` (OpenShift arbitrary-uid HOME) and `scripts/build-agent-image.mjs` (build or pull the base image; digest recorded). Add `checkDockerAvailable`/`checkDockerTmuxAvailable`/`containerApiUrl`/`hostGatewayAlias` (IS_TEST_MODE no-op) and `GET /api/docker/status`. Test: probe stub returns available/caps/Desktop flags under VITEST; `containerApiUrl` preserves scheme+port and swaps host per engine; status route returns the envelope.
- Phase 1: types + storage + schemas. Add all types (Section 3), `src/docker-hosts.ts`, `DockerHostSchema`/`DockerCaseLinkSchema`. Test: `docker-hosts.test.ts` (round-trip incl. `lastClaudeSessionId`, display path, config-hash stability); `docker-exec-options.test.ts` (schema rejects `$`/backtick in image/workdir/container).
- Phase 2: tmux-manager builders. Add `DOCKER_TMUX_SOCKET`, `dockerTmuxSessionName`, `buildDockerLaunchCommand` (resume-aware, image-check, env-prime), `buildDockerKillCommand`; wire the two ternaries + two cd-skips + Strategy 3c; harden `reconcileSessions` against docker hard-delete. Test (pure strings): adopt-proof name fails `SAFE_MUX_NAME_PATTERN`; image-check precedes create; `new-session -A` idempotent; resume flag present only when a resume id is passed; `--pull=never` present; instance label present; escaping survives `bash -c` -> `docker exec` -> `sh -lc` -> tmux WITH a host workspace path containing spaces.
- Phase 3: session.ts + mux + recovery. Add `_docker` + `resumeSessionId` threading, in-container cliVersion probe, `resolveMuxAttachCwd`, mux-interface fields, `restoreMuxSessions` passthrough, instance-scoped reaper wiring, claudeSessionId -> `DockerCase.lastClaudeSessionId` persistence, unified flag. Test: `toState()` emits docker; a persisted docker session round-trips through mux/state; a relaunch injects the persisted resume id (mock mux); reaper only targets this instance's orphaned containers.
- Phase 4: routes + first real e2e. case-routes CRUD + listing + drift-recreate; session-routes quick-start branch (scaffolding RUNS, local-availability guards skip, model accepted, effort/config rejected). Manual e2e on a real docker host: docker-host create -> docker-link -> quick-start; confirm the pane runs `claude` in the container, files land host-owned, a Codeman restart reattaches the SAME live agent, and a `docker stop` followed by relaunch RESUMES the conversation.
- Phase 5: hooks connectivity + installation. host-gateway (per engine), derived `CODEMAN_API_URL`, hook-secret mount, `CODEMAN_SESSION_ID`/`CODEMAN_MUX` exec-env + tmux setenv, host-guard allowlist, and the scaffolding write into the real workspace. Manual e2e: trigger a permission prompt from inside the container and confirm it surfaces; verify hook payloads carry the right session id. If deferred, ship docker as explicitly hook-degraded and verify output-based idle detection through the docker-exec PTY.
- Phase 6: export/import + GC + disk safety. quiesce+pause span, free-space precheck, commit+save+gzip + workspace tar + manifest + streaming download; sealed-mode refuse-or-scrub; retention/auto-prune; import with checksum validation + traversal guard + quarantined re-tag; drift-recreate; boot reaper; `runWithConversionLimit` cap; `docker rmi` in finally. Manual e2e: export, `docker load` on a second machine (or fresh case), import, confirm toolchain + workspace restored and NO creds present; attempt a sealed full-image export and confirm it is refused-or-scrubbed; attempt a `../` bundle and confirm it is rejected.
- Phase 7: frontend. Docker tab, `linkDockerCase`, run wiring, case-picker labels, panels search, caps-advisory + scaffold-warning + effort-inert notes. Verify with Playwright (`waitUntil: 'domcontentloaded'`, 3-4s settle) that the Docker tab renders and a linked docker case appears in the picker.
- Phase 8: docs + COM. Update CLAUDE.md (a "Docker cases" Key Pattern paragraph mirroring remote-SSH, plus the new state files, routes counts, and the resume/durability model), `docs/docker-cases.md`, then COM per the standard flow.
## 9. Test plan
- Unit (pure, CI-safe, mirror `test/remote-hosts.test.ts` / `test/remote-ssh-options.test.ts`):
- `test/docker-hosts.test.ts`: storage round-trip (incl. `lastClaudeSessionId`), `dockerDisplayPath`, `defaultDockerCommandForMode`, `toSessionDocker`, `containerApiUrl` (http/https, custom port, docker vs podman gateway), config-hash stability/drift, `buildDockerCreateArgs` flag ordering (cap-drop/no-new-privileges/memory==memory-swap/instance-label/`--pull=never` present; host/privileged/socket absent; per-engine uid vs `--userns=keep-id`).
- `test/docker-exec-options.test.ts`: `buildDockerLaunchCommand`/`buildDockerKillCommand` string shape and escaping through `bash -c` -> `docker exec` -> `sh -lc` -> tmux, including a workspace path with spaces; resume flag present only with a resume id; image-presence check precedes create; `dockerTmuxSessionName` fails `SAFE_MUX_NAME_PATTERN`; schema rejects `$`/backtick in image/workdir/container/name; `linkDockerCase`-shaped bodies with omitted optionals validate (no `null` on the wire).
- Probe no-op: `checkDockerAvailable`/`checkDockerTmuxAvailable`/`probeDockerCliVersion` return canned values under VITEST and never spawn.
- Integration (route tests via `app.inject()`, docker no-op'd): `/api/docker-hosts` CRUD; `/api/cases/docker-link` dup-check + broadcast; `GET /api/cases` includes the docker case with `location: 'docker'`; `/api/quick-start` docker branch rejects `envOverrides`/`effort`/config but ACCEPTS `modelOverride`, runs the workspace-scaffolding path, and constructs a session with `docker` set + seeded resume id; `DELETE /api/cases/:name` docker-unlink; export refuse-or-scrub for sealed; import traversal rejection; reaper instance-scoping (label filter). Pick a unique port only if a live-server test is added (search `const PORT =`; 3150+).
- Manual end-to-end (real docker daemon, the mandatory "always end-to-end test" gate): build the base image; link a docker case; quick-start `claude`; verify OAuth via the mounted `~/.claude`, transcript correlation (subagent/workflow watchers show the session), host-owned files, and a working permission-prompt hook; reattach after a Codeman PROCESS restart (SAME live agent); `docker stop` then relaunch and confirm conversation RESUME; reboot-equivalent (daemon restart) and confirm boot recovery recreates+resumes; change the host's memory/image and confirm the drift-recreate prompt fires; export (convenient) and confirm the tar `docker load`s with no creds; attempt a sealed full-image export and confirm refuse-or-scrub; import into a fresh case; delete the case and confirm `docker rm -f` plus instance-scoped reaper GC; confirm a docker-down state surfaces a docker-specific error and does NOT trip the generic PTY-exit breaker.
## 10. Open decisions for the user
1. Credential + blast-radius posture (combined). Convenient default bind-mounts host `~/.claude` etc. RW AND an arbitrary host workspace RW into a network-enabled container, so container-run agent code can read/modify those host trees and reach the network at the same time. Recommended: convenient default plus a per-host SEALED opt-in (`mountCredentials:false` + `network:none`) for untrusted work. Please confirm you accept the combined arbitrary-workspace-plus-egress-plus-host-creds posture for the default profile (it is still a net improvement over today's on-host skip-permissions execution).
2. Base image ownership, registry, and freshness. The `codeman/agent:base` placeholder implies a Docker Hub org the project may not own. Pick the real registry/namespace (GHCR under the repo is the natural fit), decide digest pinning, and set a REBUILD CADENCE so agents are not stuck on a stale baked `claude` (the in-container version probe surfaces staleness, but something must trigger rebuilds). Choose: pull a pinned published image, build locally on first use via `scripts/build-agent-image.mjs`, or both.
3. Container CWD strategy. Mirror the host workspace path inside the container (recommended: makes transcript projHash correlate, file features and resume capture work) vs a fixed `/workspace` (simpler mount, breaks watcher correlation). Please confirm the mirror approach.
4. Hooks in the MVP AND workspace scaffolding. Making docker hooks fire requires WRITING `.claude/settings.local.json` (and the CLAUDE.md scaffold) into the user's REAL linked host directory, a behavioral shift from "link a dir" to "link and scaffold a dir." Choose: wire hooks + scaffolding now (Phase 5, recommended, and it also enables the model picker), or ship docker as explicitly hook-degraded (no permission prompts / hook-idle) for v1 and add later. Confirm you are OK with Codeman mutating the linked host workspace.
5. Session-kill teardown and RESUME (reframed honestly). `docker stop` on session kill is not merely "free RAM vs instant reattach": it destroys the in-container live agent, and the conversation survives ONLY because the next launch runs `--resume` from the bind-mounted transcript. Choose: keep the container running (costs RAM, preserves the exact live in-flight agent) vs stop and rely on `--resume` (frees RAM, may lose uncommitted in-flight tool state). Case-delete always `docker rm -f`.
6. Rootless enforcement posture. Under rootless without cgroup-v2 systemd delegation, `--memory`/`--cpus`/`--pids-limit` are SILENTLY ignored. Choose: REQUIRE delegation (refuse to link a host that cannot enforce caps) or ship-with-warning ("resource caps are advisory on your engine"). The probe reports `capsEnforced` either way.
7. Default resume behavior. Should a re-linked or re-run docker case default to resuming its last conversation (`resumeOnStart:true`, using `DockerCase.lastClaudeSessionId`) rather than starting clean? This is the crux of making the durability story real and is the recommended default, but it changes user-visible behavior (a new session in an existing case continues the prior conversation).
8. Export defaults and disk budget. Default export button: workspace-only (fast, small, files-only, recommended for 24h+ runs) vs full-image (reproducible env, multi-GB). Also set the retention cap (max retained exports), the auto-prune policy, and the free-space threshold below which export is refused (a full `/var/lib/docker` breaks EVERY session on the host, not just docker ones).
9. Remote docker daemon (`-H ssh://...` / `--context`). Support in the MVP (composes with remote hosts, adds host-root trust surface) or local-daemon-only first.
10. Podman parity depth. Full `--userns=keep-id` plus Quadlet boot-persistence, or Docker-first with Podman as best-effort and boot-persistence via Codeman's idempotent create-if-missing only. Note the podman host alias is `host.containers.internal`, already handled per engine.
docs/docker-cases.md
+95
View File
@@ -0,0 +1,95 @@
# Docker cases
Run a case inside an **isolated Docker container** instead of directly on the host. Any number of Codeman sessions can share one container (it is scoped to the case, not the session), so a whole project lives in a sandbox with its own network, resource caps, and filesystem, and you can **export the container to move it to another machine**.
Docker mode is a **location overlay on cases**, the direct analog of [remote SSH cases](./remote-hosts.md): where a remote case runs a local tmux pane doing `ssh host` into a durable remote tmux server, a docker case runs a local tmux pane doing `docker exec -it` into a durable **in-container** tmux server. It is not a separate `SessionMode`, so `claude` / `shell` / `opencode` / `codex` / `gemini` all work inside the container.
## One-time setup: build the base image
The container needs a base image with the agent toolchain (node, the CLIs, git, tmux). Build it locally once:
```bash
node scripts/build-agent-image.mjs # builds codeman/agent:base
# options: --engine docker|podman --image <ref> --no-cache
```
The image is **secret-free**: credentials are delivered at runtime (bind mounts or `docker exec --env`), never baked in, so exports never leak them.
## Quickest path: one-click "Run in Docker"
On the **New case → Create New** tab there's a **🐳 Run in an isolated Docker container** checkbox. Checking it alone is enough: Codeman creates the case folder in `~/codeman-cases/<name>`, spins up a hardened container with sensible defaults (auto-provisioning a shared `default` host), and starts the session inside it. No host/image/network fields to fill in.
Click the checkbox's **Container settings** to optionally tweak the predefined defaults, including a **Template** picker:
| Template | Memory | CPUs | GPUs |
|----------|--------|------|------|
| Small | 2 GB | 1 | none |
| Medium (default) | 4 GB | 2 | none |
| Large | 8 GB | 4 | none |
| GPU | 8 GB | 4 | all (needs the NVIDIA container toolkit) |
**Disk is elastic** — the container's storage grows automatically as data flows in; there is no fixed cap (bounded only by host disk). Any tweaked setting creates a dedicated per-case host so it never changes the shared `default`.
## Create a docker case (full control)
App → **New case → Docker** tab:
- **Case Name** / **Workspace Path**: the workspace is a real HOST directory bind-mounted into the container at the same path. Codeman scaffolds `CLAUDE.md` + `.claude/settings.local.json` (hooks) into it, and file previews / attachments work on the real bytes.
- **Host ID**: a reusable docker host profile (image, network, resources). Reuse the same ID across cases to share settings.
- **Network**: `bridge` (internet on, default), `none` (fully isolated), or a `custom` bridge.
- **Advanced**: memory / CPU caps, **Mount host credentials** (on = your existing `~/.claude` login just works; off = a sealed sandbox you log into inside the container), **Resume last conversation on relaunch**.
Then run it like any case (Run Claude / Run Shell / …). The first launch creates the container (`codeman-case-<name>`); subsequent sessions attach to the same one.
Equivalent API:
```bash
curl -X POST localhost:3000/api/docker-hosts -d '{"id":"local","label":"Local","image":"codeman/agent:base"}'
curl -X POST localhost:3000/api/cases/docker-link -d '{"name":"sandbox","hostId":"local","hostWorkspacePath":"/home/you/projects/sandbox"}'
curl -X POST localhost:3000/api/quick-start -d '{"caseName":"sandbox","mode":"claude"}'
```
## Lifecycle
- **Reconnect after a Codeman restart** lands back in the same live agent (the in-container tmux survives).
- **Container stop / host reboot** restarts the container and **resumes** the last conversation from the bind-mounted transcript. Claude sessions launch with a pinned conversation id (`--session-id <sessionId>`, with a `--resume` fallback when the transcript already exists), and the case remembers its last conversation (`lastClaudeSessionId`), so a relaunch after the container was stopped, rebooted, or recreated continues where it left off.
- **Killing one session** only kills that session's in-container tmux session; the shared container stays up for sibling sessions.
- **Editing the docker host config** (image, memory, network, ...) is detected on the next launch: the desired config hash is compared against the container's `codeman.confighash` label, and a mismatch refuses the launch with a "config changed, recreate?" confirm. Confirming calls `POST /api/docker-cases/:name/recreate` (refused while sessions of the case are live), which removes the container so the next launch recreates it with the new config; the workspace and the conversation survive.
- **Deleting the case** `docker rm -f`s the container (the bind-mounted workspace on the host survives). An instance-scoped boot reaper removes containers whose case is gone.
## Isolation & security
Every container runs hardened: `--cap-drop ALL`, `--security-opt no-new-privileges`, non-root (`--user <hostUid>:0` so workspace files stay host-owned), `--pids-limit`, `--memory` == `--memory-swap`, `--init`. Never `--privileged`, never the docker socket. The default **convenient** profile bind-mounts host credential dirs read-write so the common login just works (creds stay on the host, never captured by `docker commit`); the **sealed** profile (`mountCredentials:false` + `network:none`) is the opt-in for genuinely untrusted work.
Rootless engines without cgroup-v2 systemd delegation cannot enforce resource caps; linking such a host warns that caps are advisory.
## Export / Import (move to another machine)
**Export** (from the Docker tab, or `POST /api/docker-cases/:name/export`): choose
- **Full image + workspace**: `docker commit` the container to an image, `docker save` it, tar the workspace, and a manifest, all into one portable `<case>-<ts>.codeman-container.tgz` (the whole toolchain, installed packages, and files). Runs in the background; you are notified when the bundle is ready.
- **Workspace only**: just the project files (fast, small).
The container is paused across the capture so the image and workspace are consistent; a full `/var/lib/docker` is guarded against with a free-space precheck; the intermediate image is always cleaned up.
**Import** (`POST /api/docker-cases/import`, or the Manage tab): copy the `.tgz` onto the new machine's `~/.codeman/docker-exports/`, then import it into a new case. The manifest and per-member SHA-256 checksums are validated, the workspace tar is extracted with a path-traversal guard, and the image is `docker load`ed and **re-tagged into a quarantined namespace** (`codeman/imported-<case>:<ts>`) so it never overwrites a local tag. The destination supplies its own credentials, so nothing secret crosses machines.
`GET /api/docker-exports` lists bundles; `GET /api/docker-exports/:filename` downloads one; `DELETE` removes one.
## Hooks require the server to be reachable from the container
In-container hooks (permission events, hook-based idle/stop/task notifications) POST to `CODEMAN_API_URL`, which is derived as `https://host.docker.internal:<port>` (`host.docker.internal` → the docker bridge gateway, e.g. `172.17.0.1`, via `--add-host …:host-gateway`). For that callback to succeed, the Codeman server must be **listening on an interface the container can reach**.
- If Codeman binds **loopback-only** (`127.0.0.1`, the default and the production systemd config), a container reaching `172.17.0.1:<port>` cannot connect, so by default **in-container hooks do not fire**. The session still works fully: idle/stop detection falls back to **output-based** detection through the `docker exec` PTY (which always works), and claude runs with `--dangerously-skip-permissions` so there are no permission prompts to forward anyway.
- **To enable in-container hooks on a loopback-only server, set `CODEMAN_DOCKER_BRIDGE_HOOKS=1`** (env). Codeman then starts a SECOND listener bound to the docker bridge gateway (`172.17.0.1`, auto-detected; override with `CODEMAN_DOCKER_BRIDGE_HOST`) that serves **only the hook endpoints** (`/api/hook-event`, `/api/status-telemetry`) and delegates them into the same secret-gated pipeline. The bridge is host-internal (containers + host, not the LAN), and every other path returns `403`, so this does not widen your network exposure. Add `Environment=CODEMAN_DOCKER_BRIDGE_HOOKS=1` to the systemd unit and restart.
- Alternatively, bind `0.0.0.0` **with `CODEMAN_PASSWORD` set** (exposes on the LAN too).
The host-gateway mapping, `CODEMAN_API_URL` derivation, host-guard allowlist, and hook-secret mount are all wired correctly; `CODEMAN_DOCKER_BRIDGE_HOOKS` closes the last gap for loopback-only servers.
## Notes & limits
- Requires Docker (or Podman) with a reachable daemon; tmux must be present in the base image (a hard prerequisite, probed at link time).
- Per-session `envOverrides` / `effort` / per-CLI config are rejected for docker cases (they do not cross into the container); configure the container via the docker host's per-mode command override instead.
- macOS Docker Desktop takes a dedicated uid path (the baked image uid; memory caps are subject to the VM ceiling).
Design + rationale: [`docker-cases-plan.md`](./docker-cases-plan.md).
docs/multi-user-plan.md
+282
View File
@@ -0,0 +1,282 @@
# Multi-User Mode: Design Plan
Status: **IMPLEMENTED on `feat/multiuser-mode`** (phases 1-5; opt-in, off by default). Target: opt-in multi-user support behind a `--multiuser` flag, with per-user case spaces and an admin panel for user management.
Shipped by phase:
- **Phase 1** (user store + mode plumbing + CLI): `src/user-store.ts` (scrypt, atomic 0600 writes, last-admin invariants, serialized read-modify-write), `src/config/multiuser.ts`, `codeman users add|passwd|list|rm`, `--multiuser` flag, bootstrap-on-first-boot. Tests: `test/user-store.test.ts`.
- **Phase 2** (multi-user auth): parallel async auth branch (`src/web/middleware/auth.ts`), `req.authUser`, per-username rate bucket, `mustChangePassword` lockbox, `GET /api/me` + `POST /api/me/password`, QR identity-bound minting, network-bind + tunnel exemptions, new error codes. Tests: `test/multiuser-auth.test.ts`.
- **Phase 3** (ownership threading): `Session.owner` at every create path + recovery mirror; `findSessionOrFail` owner check + list filtering; §6.3 permission policy (`resolveClaudeModeForUser` at all spawn sites incl. one-shots via `buildPromptArgs`; shell/launchCommand grant); per-user case spaces (`resolveCasesDir`) + owner-scoped case list + admin-only host CRUD; `workingDir` confinement; `sessionCapacityState` per-user cap. Tests: `test/ownership-scoping.test.ts`.
- **Phase 4** (event fan-out): WS owner gate; SSE per-client identity + `broadcast`/terminal-batch routing (`deriveSseHint`, fail-closed); `getLightState` per-identity filtering; file-route preview/thumbnail/history + `GET /api/search` scoping.
- **Phase 5** (admin API + frontend): `src/web/routes/admin-routes.ts` (user CRUD, one-time passwords, last-admin guards, session revoke/kill) + `src/web/admin-audit.ts`; `public/admin-ui.js` (identity boot, change-password modal + interceptor, admin Users tab). Tests: `test/admin-routes.test.ts`, `test/admin-ui.test.ts`.
Deferred follow-ups (documented, non-blocking): away-digest + subagent/workflow REST-list scoping, push-subscription identity/routing, per-user screenshot subdirs, `linked-cases.json` v2 owner field, `ScheduledRun.owner`, plan-orchestrator internal one-shot mode resolution, and a Playwright browser pass. Phase 6 (login form replacing Basic) remains out of scope.
## 1. Summary
Today Codeman is strictly single-user: one optional credential pair (`CODEMAN_USERNAME`/`CODEMAN_PASSWORD`), one shared `~/codeman-cases` folder, one global session list, and a global SSE/WS fan-out. This plan adds an opt-in **multi-user mode**:
- **Off by default.** Without the flag, behavior stays byte-identical to today (same auth path, same paths, same payloads). All new code is gated behind `isMultiUserMode()`.
- **`codeman web --multiuser`** (or `CODEMAN_MULTIUSER=1`) enables named users with individually hashed passwords stored in `~/.codeman/users.json`.
- **Each user gets their own space**: `~/codeman-users/<username>/cases/<case>` replaces the shared `~/codeman-cases` for that user. Sessions, cases, attachments, search, digests, and SSE events are scoped to their owner.
- **Admin panel** (App Settings, admin-only "Users" tab): create/delete users, change/reset passwords, enable/disable accounts, delete a user's space, see per-user live sessions and disk usage, force logout.
## 2. Threat Model (read first, be honest about this)
Multi-user mode is **workspace separation for a trusted team, NOT security isolation between mutually distrusting users**:
- Every session still runs as the **same OS account** with `claude --dangerously-skip-permissions`. Any user can ask their agent to `cat /home/<host>/codeman-users/otheruser/...`. The web layer enforces scoping; the agent layer cannot.
- **Shell sessions and custom launch commands are the bluntest holes**: `SessionMode = 'shell'` hands out a raw shell as the host account, and a cron job's `launchCommand` runs an arbitrary command; no Claude permission classifier is involved in either. These must be gated behind the same grant as bypass (section 6.3), otherwise the `auto`-mode mitigation below is theater.
- All sessions share one tmux socket (`-L codeman`), one `~/.claude` (transcripts, credentials, plan usage), one Claude subscription.
- Mitigation for stronger isolation: pair a user's cases with **Docker cases** (container per case, `docs/docker-cases.md`), or run separate Codeman instances per user (`CODEMAN_INSTANCE`, separate OS accounts). True per-user OS isolation is explicitly **out of scope** for this feature.
- Partial mitigation at the agent layer: non-admin users default to Claude's `auto` permission mode (section 6.3), whose safety classifier blocks destructive actions and credential exfiltration. That reduces, but does not eliminate, cross-user snooping; the `canBypassPermissions` grant reopens it and should be given deliberately.
This must be stated loudly in `docs/security-architecture.md`, the README section, and the admin panel UI ("Users share the host account; this separates workspaces, it does not sandbox users from each other").
Also note the flip side: multi-user mode strictly _improves_ today's network posture, because it removes the single shared password and gives every person their own revocable credential.
## 3. Activation and Mode Rules
| Condition | Behavior |
| ------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| No flag (default) | Exactly today's behavior. `users.json` is never read. Single-user auth via `CODEMAN_PASSWORD` if set. |
| `--multiuser` / `CODEMAN_MULTIUSER=1`, `users.json` has users | Multi-user auth active. `CODEMAN_PASSWORD` is ignored for login (warn if set). |
| `--multiuser`, no `users.json` (first boot) | Bootstrap: if `CODEMAN_USERNAME`/`CODEMAN_PASSWORD` are set, create that user as the initial admin and continue. Otherwise refuse to start with instructions to run `codeman users add <name> --admin`. Never start multi-user with zero users (there would be no way in). |
| `--multiuser` on a non-loopback bind | Allowed without `CODEMAN_PASSWORD`: `server.ts start()` treats "multi-user with >= 1 enabled user" as satisfying the auth requirement in the loud-warning check (wire into the existing `isLoopbackBindHost()` branch). |
| Flag later removed | Single-user mode again. Sessions/state that carry `owner` fields keep working (owner is simply ignored); user spaces remain on disk untouched. |
Plumbing: flag in `src/cli.ts` (web command), env in a new `src/config/multiuser.ts` exporting `isMultiUserMode()`. Per-instance like everything else: a beta instance (`CODEMAN_INSTANCE=beta`) has its own `users.json` via `dataPath()`.
## 4. Data Model and Disk Layout
### 4.1 `~/.codeman/users.json` (via `dataPath('users.json')`, mode 0600, atomic write: tmp + rename)
```jsonc
{
"version": 1,
"users": [
{
"username": "alice", // canonical lowercase slug
"role": "admin", // "admin" | "user"
"password": {
"algo": "scrypt", // node:crypto scrypt, no new deps
"N": 16384,
"r": 8,
"p": 1,
"salt": "<hex 32B>",
"hash": "<hex 64B>",
},
"disabled": false,
"mustChangePassword": false, // set by admin reset; gates all API access until changed
"canBypassPermissions": false, // permission-mode grant, see section 6.3; false for new users
"createdAt": 1752900000000,
"lastLoginAt": 1752900000000,
},
],
}
```
- **Username rules**: `^[a-z0-9][a-z0-9_-]{1,31}$` (it becomes a folder name), stored lowercase, unique case-insensitively. Reserve `admin`? No: any name can be admin; role is a field, not a name.
- **Hashing**: `scrypt` from `node:crypto` with per-user salt, compared via `timingSafeEqual`. Params stored per record so they can be raised later; verify tolerates old params and rehashes on next successful login.
- New module `src/user-store.ts` (mirrors the `remote-hosts.ts` / `docker-hosts.ts` pattern): `readUsers()`, `writeUsers()`, `verifyPassword()`, `createUser()`, `setPassword()`, `deleteUser()`, plus pure helpers (`isValidUsername`, `hashPassword`) that are unit-testable without IO. In-process cache with short TTL like `readSettings`, invalidated on every write; the short TTL also covers the CLI (section 10) editing `users.json` while the server runs (cross-process changes picked up within the TTL).
### 4.2 User spaces
```
~/codeman-users/
alice/
cases/
my-project/ <- same layout as today's ~/codeman-cases/<case>
bob/
cases/
```
- New helper in `route-helpers.ts`:
`resolveCasesDir(user?: AuthUser): string`
single-user mode: returns `CASES_DIR` (today's `~/codeman-cases`); multi-user: returns `join(USER_SPACES_DIR, user.username, 'cases')`, creating it lazily on first use.
- `CASES_DIR` stays exported for single-user code paths, but every route usage (see 6) switches to the resolver.
- The **user folder** (`~/codeman-users/<username>/`) is the deletion unit for "delete user + space" and leaves room for future per-user extras (uploads, exports) beside `cases/`.
- Legacy `~/codeman-cases` in multi-user mode: surfaces to admins only, as a read-only "Unassigned (legacy)" group in the case list, with an admin action `POST /api/admin/cases/assign { case, username }` that `fs.rename`s the folder into a user's space (same-filesystem move, cheap). No automatic migration.
## 5. Auth Pipeline Changes (`src/web/middleware/auth.ts`)
Keep the existing single-user branch untouched. Add a parallel multi-user branch selected once at registration time:
1. **Credential check**: Basic header parsed into `username:password`, verified against the user store (scrypt + `timingSafeEqual`). Disabled users fail closed.
2. **Cookie sessions**: same `codeman_session` cookie and `StaleExpirationMap`, but `AuthSessionRecord` gains `username` and `role`. All existing TTL/sliding/eviction logic reused. Eviction cap becomes per-user aware (evict oldest _of that user_ first) so one user cannot flush everyone's sessions by logging in 100 times.
3. **Request identity**: decorate `req.authUser = { username, role }` (Fastify decorateRequest). In single-user mode `req.authUser` is `{ username: 'admin', role: 'admin' }` when auth is on, and a synthetic admin when auth is off, so downstream code has ONE code path.
4. **Rate limiting**: keep the per-IP bucket; add a per-username failure bucket (same `StaleExpirationMap` pattern) so a botnet cannot brute-force one account across IPs, and one flaky user behind a NAT cannot lock out the rest.
5. **`mustChangePassword` gate**: when set, every API request except `GET /api/me`, `POST /api/me/password`, and static assets returns 403 with `errorCode: 'PASSWORD_CHANGE_REQUIRED'`; the frontend intercepts that code and shows the change-password modal.
6. **Password change vs Basic-auth caching**: browsers cache Basic credentials. After a password change we revoke all of that user's cookie sessions; the next request falls to Basic with stale creds, gets 401, and the browser re-prompts. Acceptable for v1; a proper login form is Phase 6 (see 15).
7. **Unchanged**: hook-secret loopback bypass (hooks authenticate the _instance_, not a user; the event maps to a session which has an owner), host guard, Origin/CSRF guard, security headers.
8. **WS upgrade identity** (`ws-routes.ts`): the global auth `onRequest` hook does run on the upgrade request (`@fastify/websocket` v11 runs hooks before the handshake; browsers send the session cookie), but the route handler itself only checks Host/Origin and never learns WHO authenticated. Multi-user: the handler reads the decorated `req.authUser` and closes 4003 unless owner or admin (section 6.4; identity plumbing lands in Phase 2, the owner check in Phase 4 once sessions have owners). Add a regression test that an upgrade with no credentials is rejected while auth is active: the handler-level Host/Origin gate alone must never be mistaken for auth.
9. **QR auth** (`/q/:code` redemption in `system-routes.ts`, minting in `tunnel-manager.ts`): today there is ONE global token, auto-rotated every 60s with a 90s grace window. A globally-rotating token cannot carry an identity (every logged-in user sees the same code), so multi-user mode replaces rotation with **on-demand minting**: an authenticated `POST /api/tunnel/qr` mints a single-use, short-TTL token bound to `req.authUser.username` (field on `QrTokenRecord`); redemption creates a cookie session for that user. Existing rate-limit buckets (`qrAuthFailures`, global `QR_RATE_LIMIT_MAX`) apply unchanged. Single-user mode keeps the rotating token.
New error codes in `src/types/api.ts`: `FORBIDDEN`, `PASSWORD_CHANGE_REQUIRED`, `USER_EXISTS`, `USER_NOT_FOUND`, `LAST_ADMIN`.
Role guard helper in `route-helpers.ts`: `requireAdmin(req, reply): boolean` used as the first line of every admin handler (403 `FORBIDDEN`), plus `requireOwnerOrAdmin(req, session)`.
## 6. Ownership Threading (the big refactor)
### 6.1 Sessions
- `Session` gains `owner?: string` (constructor option), persisted in `SessionState.owner`, included in `toState()`, round-tripped through recovery (`mux-sessions.json` entries carry it, `restoreMuxSessions` passes it back, exactly like `remote`/`docker`).
- Every session-creating path stamps the owner from `req.authUser`. Verified inventory of `new Session(...)` call sites: `POST /api/sessions` (session-routes.ts:444), `POST /api/quick-start` (:1956), `POST /api/run` one-shot (:1652), Ralph start (ralph-routes.ts:327), **cron** (cron-service.ts:352; `CronJob` gains `owner`, stamped at job create, launched as the job's owner), legacy `ScheduledRun` loop (server.ts:1603), plan generation + plan-orchestrator agents (plan-routes.ts:128, plan-orchestrator.ts:422/578; owner = requesting user), and recovery (server.ts:2225, next bullet). Two non-paths, also verified: **respawn never constructs a new Session** (it re-spawns the PTY on the same object, so `owner` survives automatically; no inheritance logic needed), and **orchestrator-loop creates no sessions** (it schedules work onto existing idle sessions via the task queue; its scoping requirement is different: it must only pick idle sessions owned by the goal's creator).
- Recovery: `owner` must ALSO be mirrored on `MuxSession` (mux-sessions.json) and read back mux-first like `remote`/`docker` (`muxSession.owner ?? savedState?.owner`, the server.ts:2246-2250 pattern), or a reboot erases ownership on the next persist.
- Every session-reading/mutating route filters: non-admin users only see and act on `session.owner === req.authUser.username`. Centralize in `findSessionOrFail` (route-helpers.ts:87; the owner check there covers the 6 route files that use it: system/session/respawn/ralph/file/plan-routes) and in the list endpoints (`GET /api/sessions`, `GET /api/sessions/unified`, `GET /api/status`). The Phase 3 audit must grep for BOTH `sessionManager.getSession` AND direct map access (`ctx.sessions.get(` / `.has(`): ws-routes and hook-event-routes reach sessions that way and bypass `findSessionOrFail`.
- Admins see everything; every session row carries `owner` so the UI can badge it.
### 6.2 Cases
- All `CASES_DIR` call sites switch to `resolveCasesDir(req.authUser)`: `case-routes.ts` (list/create/delete/CLAUDE.md scaffolding, name-collision checks, docker quickcreate), `session-routes.ts` (quick-start case resolution, the workingDir-inside-cases env-strip check), `ralph-routes.ts` (case path resolution), and `plan-routes.ts:231` (easy to miss). Case-name-to-path resolution is currently DUPLICATED (`resolveCasePath` in case-routes.ts:82 and an inline copy in quick-start, session-routes.ts:1846-1863); consolidate into one owner-aware resolver as part of this refactor instead of patching both copies.
- Registries that map case names to metadata become owner-scoped. `remote-cases.json`/`docker-cases.json` are arrays of objects, so entries simply gain `owner?: string` (absent = legacy: admin-only). `linked-cases.json` is a flat `Record<caseName, path>` with no room for a field: it needs a v2 shape (`{ "version": 2, "cases": { "<name>": { "path": "...", "owner": "..." } } }`) with read-time migration of the v1 form; it is read in two places (case-routes AND inline in quick-start), both must move to the new reader. Case names only need to be unique per user.
- **Remote hosts and Docker hosts are machine-level resources**: CRUD on `/api/docker-hosts` and remote-host endpoints becomes admin-only in multi-user mode; regular users can _use_ hosts on their own cases but not define them. (Docker containers exec as the host account; letting any user define arbitrary `docker run` args is admin-equivalent.)
- Case deletion, exports (`docker-exports/`), and imports check ownership; export filenames get an owner prefix to avoid collisions (fits the existing `^[a-zA-Z0-9._-]+\.tgz$` download guard).
- **Workspace confinement for non-admins (the linchpin, do not skip)**: today `POST /api/sessions` accepts ANY host directory as `workingDir` (the only check is `statSync().isDirectory()`, session-routes.ts:305-318), and file-routes/attachments confine reads to `session.workingDir`. Without a new rule the whole scoping story is circular: a user points a session at `~/codeman-users/bob` (or `/home`) and the web layer itself serves that subtree, no agent needed. Rule: in multi-user mode a non-admin's `workingDir` must realpath-resolve inside their own space, enforced at `POST /api/sessions`, `POST /api/run`, cron job create AND fire time (the dir can change owners between the two), and Ralph auto-configure. Admins are unrestricted. This one rule is what makes the section 6.4 file-route line ("own space or own sessions' workingDirs") meaningful.
### 6.3 Per-user Claude permission-mode policy
Codeman now ships a global **Startup Mode** picker (App Settings, Claude CLI tab: `settings.claudeMode`, values `dangerously-skip-permissions` (default) | `auto` | `normal` | `allowedTools`; `auto` emits `--permission-mode auto`, Anthropic's classifier-guarded low-prompt mode). Multi-user mode layers a per-user policy on top of it:
- **Default for regular users: `auto` only.** A non-admin's Claude sessions are forced to `--permission-mode auto` regardless of the global `claudeMode` setting. `normal` and `allowedTools` are also permitted (they are strictly more restrictive than auto), but `dangerously-skip-permissions` is NOT.
- **Bypass is an explicit admin grant**: `canBypassPermissions: true` on the user record (default `false`, section 4.1). Only with that grant does the global skip-permissions default (or a future per-user choice) apply to their sessions.
- **Admins** are unrestricted; the global setting applies to them as-is.
- **Single enforcement point**: a pure `resolveClaudeModeForUser(globalMode, user)` in `user-store.ts`, applied server-side at option-resolution time, BEFORE the Session constructor, so both downstream arg builders inherit it for free (`buildPermissionArgs` in session-cli-builder.ts for the direct-PTY path AND `buildClaudePermissionFlags` in tmux-manager.ts for tmux panes; there are two builders, not one). Call sites where `getClaudeModeConfig()` feeds a spawn: session-routes.ts:452/1964, ralph-routes.ts:334, cron-service.ts:360, and recovery (server.ts:2214/2233). Recovery re-reads the GLOBAL setting on reboot, so the resolver must run there with the RECOVERED owner, or a restart silently un-downgrades every restored session. Never resolved in the frontend, so it cannot be bypassed via payload.
- **Downgrade, don't error**: a non-granted user whose effective mode would be bypass gets `auto` silently (logged + surfaced as a badge on the session), so shared presets keep working.
- **Other CLIs' bypass equivalents** follow the same grant: Codex `--dangerously-bypass-approvals-and-sandbox` (`codexDangerouslyBypassApprovals`) and Gemini `--approval-mode yolo` are refused for non-granted users (Gemini falls back to `auto_edit`, Codex to its default sandbox). Whether this stays one grant or splits per-CLI is an open question (section 15).
- **Shell mode and custom launch commands follow the grant too**: `mode: 'shell'` sessions and cron `launchCommand` are arbitrary command execution as the host account, strictly stronger than any bypass flag, and no permission-mode downgrade applies to them. Non-granted users get 403 `FORBIDDEN` on shell session/quick-start creation and on cron jobs carrying `launchCommand` (checked at create AND at fire time). Folding them under `canBypassPermissions` keeps the model one-bit; section 15 asks whether it should split.
- **Admin UI**: a "Can skip permissions" toggle per user in the Users tab (PATCH field, section 8), with a warning echoing the section 2 threat model.
- Revoking the grant takes effect on the user's NEXT session start; live sessions are listed so the admin can restart them.
### 6.4 Everything else that lists or streams
| Surface | Scoping rule |
| -------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| SSE `/api/events` | Per-connection filter (see 7) |
| WS terminal (`ws-routes.ts`) | Handler reads `req.authUser` (section 5.8) and closes 4003 unless owner or admin; today it checks Host/Origin only and has no identity |
| `GET /api/search` | `harvestSources()` only over owned sessions |
| `GET /api/away-digest` | Aggregate only owned sessions/events |
| `GET /api/subagents`, workflow runs | Filter by owning session (`claudeSessionId -> session -> owner`); agents not attributable to any session: admin-only |
| Push (`push-routes.ts`) | Subscription records currently carry NO identity (keyed by endpoint only): `subscribe` stamps `username`. All 8 `PUSH_EVENT_MAP` events are session-scoped, so routing = resolve owner from `data.sessionId`, deliver to that owner's (plus admins') subscriptions. Legacy identity-less subscriptions: admin-only delivery |
| Screenshots `/api/screenshots` | Per-user subdir `~/.codeman/screenshots/<username>/` in multi-user mode. Note: `GET /:name` deliberately rejects `/` in names as traversal, so derive the subdir server-side from `req.authUser` and keep client-visible names flat |
| Attachments | Already session-scoped; inherits the session owner check. `attachmentConfineToWorkspace` is a global, default-OFF setting today: in multi-user mode it is FORCED ON for non-admins regardless of the setting (their attachments must resolve inside their own space); the setting keeps meaning what it means for admins |
| File routes (browse/preview) | Path allowlist adds: non-admin paths must resolve (realpath) inside their own space or their own sessions' workingDirs |
| Settings (`settings.json`) | Global, admin-only writes in multi-user mode; reads allowed (per-device display keys stay in localStorage as today). Per-user server settings: out of scope v1 |
| System ops (self-update, tunnel toggle, span-displays, docker image build) | Admin-only |
| `getLightState` init snapshot | Filtered per connection. Actual contents to filter (verified): `sessions`, `scheduledRuns`, `respawnStatus`, `subagents`, `workflowRuns`, `planUsage` (host-plan telemetry: admin-only); `globalStats` stays coarse-global. Cron jobs are NOT in the snapshot (they have their own REST route; filter there). The snapshot is cached process-wide (`LIGHT_STATE_CACHE_TTL_MS`): either key the cache per role/user or filter AFTER the cache on each send |
## 7. SSE Event Filtering
`/api/events` currently broadcasts everything to everyone. Ground truth first (verified): `broadcast()` lives in `SseStreamManager` (`sse-stream-manager.ts`), not server.ts; clients are keyed by the raw Fastify reply (`sseClients: Map<FastifyReply, Set<string> | null>`, plus `sseClientsById` for live filter updates); the existing `?sessions=` filter is a bandwidth optimization applied ONLY to `session:terminal` batches in `flushSessionTerminalBatch()`, while `broadcast()` itself loops ALL clients unconditionally. The single-client delivery primitive already exists (`sendSSE`, used for the per-connection init snapshot). Plan:
- At connection time, resolve `req.authUser` and store `{ username, role }` with the client. Concretely: extend `addClient(reply, sessionFilter, isRemote, clientId)` to take the identity and change the `sseClients` map value to `{ filter, identity }` (or add a parallel `Map<reply, identity>`); there is no per-client record object today to hang it on.
- `broadcast()` gains an optional routing hint: `broadcast(event, data, { sessionId?, adminOnly?, username? })`. Resolution order per client: admin sees all; `username` targets one user; `sessionId` resolves owner via SessionManager; `adminOnly` for machine-level events (docker image builds, tunnel, self-update); no hint = broadcast to all (connection status etc.).
- **Enforce the identity check in BOTH `broadcast()` AND `flushSessionTerminalBatch()`**: the terminal batch path does not go through `broadcast()`, and it carries the highest-value payload (raw terminal bytes).
- Sweep of the ~120 backend event constants in `sse-events.ts`: mechanically, everything `session:*`, `ralph:*`, `respawn:*`, `subagent:*`, `workflow:*`, `attachment:*`, `cron:*` (job owner) carries or can resolve a sessionId/owner; `docker:*`, `system:*`, tunnel and update events are adminOnly; a short tail needs case-by-case decisions during implementation.
- The existing `?sessions=` filter and `/api/events/subscribe` compose with (never override) the ownership filter: the subscription filter can only narrow within what the identity allows.
## 8. Admin API (`src/web/routes/admin-routes.ts`, new module + `AdminPort`)
All handlers: multi-user mode only (404 otherwise), `requireAdmin`, Zod schemas in `schemas.ts`, `ApiResponse` envelope, audit-logged.
| Endpoint | Behavior |
| ------------------------------------------------ | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `GET /api/admin/users` | List users + stats: role, disabled, createdAt, lastLoginAt, live session count, case count, space disk usage (best-effort async walk, cached 60s), active cookie-session count |
| `POST /api/admin/users` | Create: `{ username, role, password? }`. No password given: generate a one-time password, return it ONCE in the response, set `mustChangePassword` |
| `PATCH /api/admin/users/:username` | `{ role?, disabled?, canBypassPermissions? }`. Demoting/disabling the last enabled admin: 409 `LAST_ADMIN`. Disable also revokes cookie sessions. `canBypassPermissions` is the section 6.3 grant (default false) |
| `POST /api/admin/users/:username/reset-password` | Generates one-time password (returned once), sets `mustChangePassword`, revokes cookie sessions |
| `POST /api/admin/users/:username/logout` | Revoke all cookie sessions for that user. Honest limit under Basic auth: the browser silently re-sends cached credentials and gets a fresh cookie on the next request, so logout only truly ends QR-issued sessions; to actually lock someone out, disable the account or reset the password. Say so in the panel tooltip until Phase 6 |
| `DELETE /api/admin/users/:username` | `{ deleteSpace?: boolean }` (default false). Refuses last admin. Kills the user's live sessions first (normal kill flow, incl. docker/remote teardown per case), revokes cookies, removes from store. With `deleteSpace`: guarded recursive delete of `~/codeman-users/<username>` (realpath must be inside `USER_SPACES_DIR`, top-level dir must not be a symlink), plus their registry entries and push subscriptions |
| `POST /api/admin/cases/assign` | Move a legacy `~/codeman-cases/<case>` into a user's space (`fs.rename`) |
| Self-service `GET /api/me` | `{ username, role, mustChangePassword }` (works in single-user mode too: synthetic admin; the frontend uses it to decide whether to render admin UI) |
| Self-service `POST /api/me/password` | `{ currentPassword, newPassword }`, verifies current, min length 8, revokes other sessions, clears `mustChangePassword` |
**Audit log**: append-only `~/.codeman/admin-audit.jsonl` (same idiom as `session-lifecycle.jsonl`): timestamp, acting admin, action, target, request IP. User management without an audit trail is not acceptable even for a homelab tool.
SSE additions (both `sse-events.ts` and `constants.js`): `admin:usersChanged` (adminOnly; the panel re-fetches) and `auth:passwordChangeRequired` (targeted to the user).
## 9. Frontend
- **`GET /api/me` on boot** (app.js init): stores `window.__codemanUser`; everything below keys off it. Single-user mode returns the synthetic admin, so the UI needs no mode awareness beyond "am I admin".
- **Admin panel**: new tab "Users" in the App Settings modal (settings-ui.js), rendered only for admins in multi-user mode. Table of users with actions (create, reset password showing the one-time password in a copy-to-clipboard reveal, enable/disable, role toggle, logout, delete with a typed-username confirm for the delete-space variant). No new header button (mobile header policy test stays green; the settings modal is already reachable everywhere).
- **Change-password modal**: shown on `PASSWORD_CHANGE_REQUIRED` (fetch interceptor in api-client.js) and reachable from settings for self-service.
- **Owner badges**: admin's session tabs and the session palette/manager show `owner` on foreign sessions; regular users see no change.
- New module `admin-ui.js` if the settings-ui.js addition gets large (load order after settings-ui, before session-ui), else keep inside settings-ui.js. Follow the `@fileoverview` + `@loadorder` convention either way.
## 10. CLI Additions (`src/cli.ts`)
Headless bootstrap and recovery must not require the web UI:
```
codeman users add <name> [--admin] # prompts for password (hidden input), or --password-stdin
codeman users passwd <name> # reset password
codeman users list
codeman users rm <name> [--delete-space]
```
These operate directly on `users.json` via `user-store.ts` (no server needed), honoring `CODEMAN_INSTANCE`. This is also the answer to "locked out: last admin forgot password".
## 11. Limits and Config
- New `src/config/multiuser.ts`: `isMultiUserMode()`, `USER_SPACES_DIR` (`~/codeman-users`, overridable via `CODEMAN_USER_SPACES_DIR` for tests), `MAX_USERS` (default 25), per-user session cap (default: global cap / 2, env `CODEMAN_MAX_SESSIONS_PER_USER`).
- Cap enforcement is currently COPY-PASTED: the global `MAX_CONCURRENT_SESSIONS` (50, `config/map-limits.ts:25`) check appears at 6 independent sites (session-routes.ts:298/1622/1683, ralph-routes.ts:275, cron-service.ts:340, server.ts:1595). Do not add a 7th copy per site: extract one `assertSessionCapacity(ctx, owner?)` helper doing the global + per-user checks and use it everywhere, or the per-user cap WILL miss a path.
- Global limits (50 sessions, SSE clients 100, terminal buffers) are unchanged and shared; the per-user session cap is the fairness lever.
## 12. Compatibility Matrix
| Concern | Guarantee |
| ------------------ | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Default (no flag) | No behavior change. No new file reads on the hot path. All new fields optional in state |
| State round-trip | `SessionState.owner`, `MuxSession.owner`, `CronJob.owner`, registry `owner` fields are optional; old state loads clean; new state loaded by an old build ignores unknown fields (existing tolerant parsing) |
| Instance isolation | `users.json`, audit log, screenshots subdirs all via `dataPath()`; user spaces dir is shared across instances like `~/codeman-cases` is today (documented) |
| API versioning | HTTP API is internal per `docs/versioning-policy.md`; still, all changes are additive. Ship as a **minor** version |
| Hooks | Unchanged (instance-level hook secret; owner resolved from the session) |
## 13. Implementation Phases
Each phase is independently shippable behind the flag and ends with its tests green.
**Phase 1: user store + mode plumbing** (no behavior change yet)
`src/user-store.ts`, `src/config/multiuser.ts`, CLI `users` subcommands, bootstrap-on-first-boot logic, `users.json` schema + atomic writes.
Tests: `test/user-store.test.ts` (hashing, verify, params upgrade, username validation, atomic write, last-admin invariants; pure, no server).
**Phase 2: multi-user auth**
Auth middleware branch, `req.authUser` decoration, cookie records with username/role, per-username rate bucket, `mustChangePassword` gate, WS upgrade identity plumbing + unauthenticated-upgrade regression test (section 5.8), QR on-demand minting + identity binding (section 5.9), `GET /api/me`, `POST /api/me/password`, error codes, network-bind check integration.
Tests: `test/multiuser-auth.test.ts` (live server, unique port 3170+; wrong password, disabled user, cookie carries identity, per-user rate limit isolation, mustChangePassword lockbox, QR redemption identity). Reuse the `delete process.env.CODEMAN_PASSWORD` idiom from `test/setup.ts`.
**Phase 3: ownership threading**
Session `owner` + persistence + `MuxSession` mirror + recovery; `resolveCasesDir()` refactor across case/session/ralph/plan routes (consolidating the duplicated case-path resolution); registry owner fields incl. the linked-cases v2 shape; `findSessionOrFail` owner check + the direct-`sessions.get` audit; list filtering; owner stamping across ALL create paths from 6.1; **non-admin workingDir confinement** (6.2); permission-mode/shell/launchCommand policy (6.3); `assertSessionCapacity` helper + per-user cap.
Tests: `test/routes/ownership-scoping.test.ts` (inject-based: user A cannot read/kill/input user B's session, case lists are disjoint, admin sees both), extend `test/cron-service.test.ts` for owner stamping, recovery round-trip in the existing mux-recovery tests.
**Phase 4: event fan-out + remaining surfaces**
SSE routing hints + client identity (enforced in BOTH `broadcast()` and the terminal-batch flush), WS owner gate (identity landed in Phase 2), search/digest/subagent/workflow scoping, push subscription identity + owner routing, screenshot subdirs, file-route scoping, `getLightState` filtering + per-identity caching, admin-only system ops.
Tests: `test/sse-ownership.test.ts` (two SSE clients, event for A's session reaches only A + admin), WS upgrade rejection test, search/digest scoping tests.
**Phase 5: admin API + frontend**
`admin-routes.ts` + `AdminPort` + schemas + audit log + `admin:usersChanged`; settings-ui Users tab, change-password modal, owner badges, api-client interceptor.
Tests: `test/routes/admin-routes.test.ts` (CRUD, last-admin 409, one-time password flow, delete-space guard rails incl. symlink refusal), frontend vm-sandbox test following `test/run-mode-ui.test.ts` pattern, Playwright pass per the always-end-to-end rule before calling it done.
**Phase 6 (optional, later): login page**
Replace Basic with a form + `POST /api/login` in multi-user mode only (fixes browser credential caching UX, enables logout button). Explicitly deferred; Basic works for v1.
**Docs**: update `docs/security-architecture.md` (new section: multi-user model + threat model from section 2), `README.md` (short opt-in section), `CLAUDE.md` (Key Patterns entry + State Files + route/SSE counts), this file gets a "shipped" status stamp per phase.
## 14. Key Risks / Decisions Made
1. **Not a security boundary at the agent layer** (section 2). Decided: ship with loud documentation; Docker cases are the isolation story.
2. **`findSessionOrFail` as the single enforcement point** for ~30 session routes: any route that fetches sessions another way must be audited in Phase 3 (grep for `sessionManager.getSession` outside route-helpers).
3. **SSE sweep is the riskiest surface**: a missed event leaks metadata (not terminal content, which is session-scoped, but names/paths). Phase 4 includes a checklist pass over all ~138 events with the default flipped to "owner-scoped unless explicitly global": fail closed.
4. **Basic-auth password-change UX** is mediocre (browser re-prompt). Accepted for v1; Phase 6 fixes it properly.
5. **Legacy case migration** is manual (admin assigns). No silent moves of user data.
6. **Case-name uniqueness becomes per-user**; tmux session names already include the session id so no collision, but the `w<n>-<case>` tab naming and lifecycle-log rows should include the owner for disambiguation in admin views.
7. **`workingDir` confinement (6.2) is the single most load-bearing rule**: every file-serving and agent-spawning surface downstream trusts `session.workingDir`. Review and test it as carefully as the auth branch (foreign-space path, symlink into a foreign space, `..` traversal, cron fire-time re-check).
8. **The WS handler never sees identity today** (auth happens only in the global hook): the 5.8 wiring is new code on a security-sensitive path; cover unauthenticated, foreign-user, and admin upgrades with tests.
## 15. Open Questions (answer before Phase 3)
1. Should admins' own cases live in `~/codeman-users/<admin>/cases` (symmetric, proposed) or keep using legacy `~/codeman-cases`? Proposed: symmetric; legacy dir is a migration source only.
2. Per-user settings (respawn presets, notification prefs): global-only in v1. Worth a `users/<name>/settings.json` overlay later?
3. Should regular users be allowed to create Docker cases on admin-defined hosts (proposed: yes) or is Docker entirely admin-only?
4. Session handoff: does an admin need "reassign session/case to another user"? (Cheap to add next to `cases/assign`; not in v1 scope.)
5. Permission-mode grants (section 6.3): one `canBypassPermissions` flag covering Claude/Codex/Gemini bypass equivalents PLUS shell mode and cron `launchCommand` (proposed: one flag, keep it one-bit), or split into `canBypassPermissions` + `canRunArbitraryCommands`? And should admins be able to set a per-user DEFAULT mode (for example force `normal` for an intern) rather than just gating bypass?
6. OpenCode has no single bypass flag (its permission config rides `OPENCODE_CONFIG_CONTENT`): decide what the grant means there before Phase 3, or exclude OpenCode mode for non-granted users in v1.
docs/security-architecture.md
+33 -2
View File
@@ -30,7 +30,8 @@ an explicit, guided opt‑in.
7. [Supply‑chain & build‑asset hardening](#7-supplychain--buildasset-hardening-cod28)
8. [Multi‑instance isolation](#8-multiinstance-isolation)
9. [Transport security headers](#9-transport-security-headers)
10. [Quick reference](#10-quick-reference)
10. [Docker container isolation](#10-docker-container-isolation)
11. [Quick reference](#11-quick-reference)
---
@@ -471,7 +472,35 @@ production layout (`~/.codeman`, `-L codeman`, port 3000).
---
## 10. Quick reference
## 10. Docker container isolation
Docker cases (1.4.0) run a session inside a per‑case container instead of on the host. The security posture:
- **Hardened create flags, always** — `--cap-drop ALL`, `--security-opt no-new-privileges`, `--pids-limit` (fork‑bomb guard), `--memory` == `--memory-swap` (a real OOM cap), `--init`, and non‑root: `--user <hostUid>:0` on Linux (host uid → workspace files stay host‑owned; GID 0 keeps `$HOME` writable), `--userns=keep-id` on rootless Podman. **Never** `--privileged`, and **never** the docker socket — the pure builder in `docker-hosts.ts` cannot emit them and the schema cannot represent them.
- **Credentials never enter an image** — the convenient default bind‑mounts host cred dirs (`~/.claude`, `~/.codex`, `~/.gemini`, `~/.config/{gcloud,opencode}`) read‑write. Bind mounts are physically excluded from `docker commit`, so exported images are secret‑free. API‑key CLIs get their key as an exec‑time NAME‑ONLY `--env OPENAI_API_KEY` (no `=value`, no `ps` leak, never committed); a create‑time `-e` for a secret is never used. The **sealed** profile (`mountCredentials:false` + `network:none`) drops the host mounts; full‑image export is then refused (an in‑container login would ride the committed layer) unless a pre‑commit scrub is opted into.
- **Blast radius — accept it explicitly** — the convenient profile mounts an arbitrary host workspace RW plus the host credential dirs RW into a network‑enabled container, so container‑run agent code can read/modify those host trees and reach the network at once. Still a net improvement over today's on‑host `--dangerously-skip-permissions` execution; use the sealed profile for genuinely untrusted work.
- **Import is untrusted‑bundle‑safe** — `/api/docker-cases/import` validates the manifest + per‑member SHA‑256 before extraction, rejects absolute / `..` tar members (traversal guard), and re‑tags the loaded image into a quarantined namespace so it can never overwrite `codeman/agent:base` or a pre‑existing tag.
- **Host guard & the bridge‑hooks listener** — in‑container hook callbacks carry `Host: host.docker.internal` / `host.containers.internal`; both are on the always‑on host‑header allowlist (`DOCKER_HOST_GATEWAY_ALIASES`) and resolve to the host only from inside a container netns, so they are not a browser DNS‑rebinding surface. On a loopback‑only server, in‑container hooks are opt‑in via `CODEMAN_DOCKER_BRIDGE_HOOKS=1`, which binds a SECOND listener on the docker bridge gateway serving **only** the hook endpoints (every other path → `403`) into the same hook‑secret‑gated pipeline. The bridge is host‑internal (containers + host), not the LAN, so it does not widen network exposure; the hook secret is bind‑mounted read‑only and referenced by path.
- **Instance isolation** — every managed container is labeled `codeman.instance=<CODEMAN_INSTANCE>`; the boot reaper reaps orphans of its OWN instance only, so a beta never removes a prod container. The in‑container tmux socket (`-L codeman-docker`) + session name (`codeman-dkr-*`) deliberately fail a nested Codeman's discovery pattern.
Full feature guide: [`docker-cases.md`](docker-cases.md).
---
## 10a. Multi‑user mode (opt‑in)
`codeman web --multiuser` (or `CODEMAN_MULTIUSER=1`) turns on named users with individually scrypt‑hashed passwords in `~/.codeman/users.json` (mode 0600). OFF by default; when off, nothing here applies and behavior is byte‑identical to single‑user. Design + phase status: [`multi-user-plan.md`](multi-user-plan.md).
- **It is workspace separation, NOT a security boundary between users.** Every session still runs as the SAME OS account with agent code that can read the whole host. Any user can ask their agent to `cat` another user's files; the WEB layer enforces scoping, the AGENT layer cannot. Mitigations: give non‑admins the default `auto` permission mode (classifier‑guarded), pair users with **Docker cases** (container per case) for real isolation, or run separate Codeman instances under separate OS accounts. Stated loudly in the admin panel and the plan's threat model (section 2).
- **It strictly improves network posture.** It removes the single shared `CODEMAN_PASSWORD` and gives each person a revocable credential; a non‑loopback bind and the tunnel‑enable guard are satisfied by "multi‑user with ≥1 enabled user" without a shared password.
- **Auth is a parallel branch** (`middleware/auth.ts`) that leaves the single‑user path untouched: per‑user scrypt verify (`timingSafeEqual`, timing‑equalized against user enumeration), identity‑carrying cookies, a per‑username failure bucket (a botnet can't brute one account across IPs; one NATed user can't lock out the rest), and a `mustChangePassword` lockbox. The hook‑secret loopback bypass, host guard, and Origin/CSRF guard are unchanged (hooks authenticate the INSTANCE, not a user).
- **Ownership is enforced server‑side only** and fails closed: `req.authUser` (a synthetic admin in single‑user), `findSessionOrFail` returns NOT_FOUND (never 403) for a foreign session, list/SSE/WS/file‑preview/search all filter by `session.owner`, and SSE routing defaults session‑scoped events to their owner (unresolved owner → withheld). The load‑bearing rule is **non‑admin `workingDir` confinement**: a non‑admin's session/one‑shot working dir must realpath‑resolve inside `~/codeman-users/<name>/cases`, checked BEFORE any disk write.
- **Privileged actions are a one‑bit grant** (`canBypassPermissions`, default off): only granted users (and admins) get `--dangerously-skip-permissions` (others are silently downgraded to `--permission-mode auto`), shell‑mode sessions, cron `launchCommand`, and other CLIs' bypass flags. Machine‑level resources (remote/Docker host definitions, tunnel, self‑update, settings writes) are admin‑only.
- **Admin actions are audited** append‑only to `~/.codeman/admin-audit.jsonl` (acting admin, action, target, IP). Passwords set by an admin create/reset are one‑time (returned once, force change). Under Basic auth, `logout` only truly ends QR‑issued sessions — to lock someone out, disable the account or reset the password (a proper login form is a deferred Phase 6).
---
## 11. Quick reference
| Env / flag | Effect |
|------------|--------|
@@ -482,6 +511,8 @@ production layout (`~/.codeman`, `-L codeman`, port 3000).
| `--https` | Enable TLS (adds HSTS) |
| `CODEMAN_INSTANCE` | Scope tmux socket + data dir for isolation |
| `CODEMAN_GESTURE=1` | Make the gesture overlay available (widens CSP) |
| `CODEMAN_DOCKER_BRIDGE_HOOKS=1` | Serve the hook endpoints on the docker bridge gateway (host‑internal, hooks‑only, `403` elsewhere) so in‑container hooks reach a loopback‑bound server — see §10 |
| `CODEMAN_DOCKER_BRIDGE_HOST` | Override the bridge gateway IP the hooks listener binds (default: auto‑detect) |
**Audit log:** session lifecycle and server start are recorded in
`~/.codeman/session-lifecycle.jsonl`.