Repository navigation
Description
T27 OS should be a working home for the coding agents people already use (Claude Code, DSH, Codex, Gemini CLI) without a separate interface per coder. This issue is the first slice: one .t27 contract that says, per coder and per transport, what is present, what is wired and what is unsupported, plus the session, errand, permission and recovery rules that every coder shares. It reads the existing pane layer; it changes no host file. Pane-driving defects (paste without submit, double errand, old scrollback, ambiguous states) stay in #13068; this issue only consumes the errand states that issue defines. Parent context: see #12945.
What was measured (read-only, 2026-11-01, this Mac)
Installed identities and their own --help:
| Coder | Installed | Interface the help shows | Not shown |
|---|---|---|---|
| Claude Code | 2.1.310 | --resume <session-id>, --session-id <uuid>, --fork-session, with --print: --output-format stream-json, --input-format stream-json, --include-partial-messages, --permission-prompts host|none |
no ACP flag |
| DSH | 0.1.0-rc.8 | --profile, commands web and plugin; dsh --profile acp fails: profile "acp" does not exist, create it with dsh plugin --profile acp add <package> |
no ACP profile installed; the documented ACP package is not on this machine |
| Codex | 0.1.2504301751 (old) | codex [options] <prompt>, -q, --approval-mode, -v <rollout> |
no exec, app-server, resume, --json, ACP |
| Gemini CLI | 0.65.0 | --acp (agent in ACP mode), -o stream-json, --session-id, --list-sessions, -r latest|index, --approval-mode |
resume takes an index, not a session id |
Consequences: the DSH ACP route is unavailable until the owner confirms an install or an upgrade; the resume-by-id assumption for Codex in specs/t27os/agents.t27 (RESUME_KIND) does not hold for the installed Codex; the comment at specs/t27os/keys.t27 lines 960-962 says Codex and Gemini are not installed, which is stale (both binaries exist, neither is a pane profile).
What T27 OS already does for Claude Code (audit)
| Aspect | Status |
|---|---|
| Launch, PTY, input, output | implemented and wired (pty_logic, pty_server); one dated execution receipt (docs run-local, 2026-10-20) for a shell pane and a live TUI in general |
| Resize | implemented in the specs; no execution receipt |
| Working directory | wired (T27OS_CWD); virtual branch or TriGit binding: not implemented |
| Send a message | wired (write verb); acceptance of the message: not implemented (see #13068) |
| Streaming, tools, permission prompts | terminal only; the app sees hook-derived state bytes, not structured events |
| History, continuation, recovery | implemented and wired (session store, --resume <sid>); no execution receipt for restart-resume |
| Termination, cancel, child processes | kill wired; cancel only by a raw key; child handling not verified |
| Subagent display | terminal only |
| Installed version identity | not in code; measured by hand above |
Events a native chat or graph can use today: pane lifecycle, hook-derived state, session id and model, key and profile and failover count, lanes table, git branch, the live events ring. Events that stay inside the terminal: conversation text, tool calls and diffs, permission prompt text and options, subagent output, API error rows, token and cost usage.
Boundary
- specs/t27os/coder_registry.t27
User Scenarios
- Given a pane for a coder is about to be offered, when the registry is asked which routes it has, then each route is reported at one evidence level (none, described, present, wired, verified) and an unsupported capability is reported as unsupported, never inferred from the coder name.
- Given an errand was sent and the app restarted before the result was read, when the session is recovered, then the registry refuses to resend until the result is reconciled.
- Given a coder asks for a permission, when the answer comes from a neighbouring pane or an unrecorded source, then the registry refuses it.
Requirements
- FR-001: The contract MUST define provider ids for Claude Code, DSH, Codex and Gemini CLI, and the transports PTY, ACP and JSON stream, with an evidence level per provider and transport pinned by dated tests.
- FR-002: The contract MUST NOT mark a capability usable below the wired level, and MUST report PTY as reaching only the state event.
- FR-003: The contract MUST define the errand lifecycle (none, sent, accepted, running, returned, failed, unknown) and MUST refuse a resend of a returned or unknown errand.
- FR-004: The contract MUST define the permission flow so that only a user answer or a recorded policy answers a prompt, and a grant applies to its own pane only.
- FR-005: The contract MUST define recovery after disconnect (reattach, resume, fresh, refuse) from the session identity kind, and MUST state that a view switch is a new session where the coder does not support one session in both modes.
- FR-006: The contract MUST bind a writing session to a workspace that is allowed and isolated, using the existing
dsh_scoperoot check, and MUST refuse a second writer on an unisolated workspace. - FR-007: The contract MUST import existing helpers (agents table, DSH bridge method names, scope check, a string-equal helper) and MUST NOT copy a body that exists elsewhere.
- FR-008: The contract MUST carry no key, token or credential value.
Success Criteria
t27c check specs/t27os/coder_registry.t27exits 0 with 0 errors.t27c test-report specs/t27os/coder_registry.t27exits 0 with FAIL 0 and 0 vacuous passes, run in a T27 OS pane (the Mac gate hook forbids it here).git diff --stat masterlists 1 file underspecs/.
Activity
gHashTag commentedon Nov 17, 2026
Spec created: specs/t27os/coder_registry_contract.t27 (typecheck OK, 1 test pass).
gHashTag commentedon Nov 17, 2026
Reopened: closed as spec-only. A .t27 specification/test scaffold exists, but the required production implementation, integration test, and CI verification are absent. This issue remains open until the target component implements the specified behavior and evidence is provided.
gHashTag commentedon Nov 17, 2026
Specs created for M10 milestone. Typecheck OK.
gHashTag commentedon Nov 18, 2026
Reopened: closed as spec-only. A .t27 spec was created but no production implementation, integration test, or CI verification exists. This issue remains open until the target component implements the specified behavior and evidence is provided.