Repository navigation
Build 5/7 — Session naming: official names, immutable labels, self-report only for unnamed Codex #78
Description
Part of #67 · Build group 5 of 7 from the signed port table (amended 2026-08-25: Codex self-report route dropped). Revised after the Codex consistency review.
Rows covered
P10 Stable Session naming — adapt Claude to the official name; adapt Codex to daemon Thread.name with a <project> · <short thread id> display fallback.
Legacy citations (verbatim from the table)
legacy@1d32845:bridge/hook.py:215-253; legacy@1d32845:bridge/labels.py:20-70,73-106 (<project> · <title>, Git common-dir project name, immutable once accepted, never a target); legacy@1d32845:bridge/daemon.py:1504-1544; legacy@1d32845:bridge/store.py:1875-1902 (left behind). Official sources: claude agents --json name; Codex Thread.name (app-server-protocol/src/protocol/v2/thread.rs:2185-2255 @3a469a2).
Shape
seams/identity.pySessionLabelstays the value (:1-15,28-82already rejects empty/malformed halves). A label is never a target (seams/identity.py:3-10); targetability depends only onreachandchild(Build 1/7 — Discovery, roster and reachability: lock the one Session-inspection type #74) — never on whether a name exists.adapters/agent/_naming.py(shared private module), ≤ 100 lines, pure:def compose(project_name: str, title: str) -> SessionLabelwith legacy's title validation (non-empty, single line). The I/O boundary is the caller: each lane adapter resolvesproject_namefrom the Session'sworkspaceviagit rev-parse --git-common-dir(aslegacy labels.py:20-70) inside itsinspect, and passes the string in.- Claude:
title= the rostername. Codex attached:title= daemonThread.namewhen present, else<short thread id>(first 8 chars) — composed locally, nothing asked of the Session. Unattached rows are labelled the same way from their provenance facts; they are still not targets, for reach reasons only. - Core applies a label exactly once per exact
SessionTargetfor the row's lifetime. Two rows with an identical label in one project are not disambiguated by Core:match_label()(core/sessions.py:120-150) keeps refusing ambiguity and the user is asked; the roster payload may showfirst_seenorder for display only. - Children (Build 6/7 — Child-process rule: seen in the roster, never spoken to, never announced #79) get no invented name — table evidence risk 1.
Left behind
Transcript-derived ai-title, mutable label history, durable label rows, the legacy self-report ingress (hook.py:215-253, daemon.py:1504-1544 — cited as behaviour evidence for immutability only).
Red first
Acceptance step stable name red before, green after: the same name across three consecutive status calls and across a Stop, both lanes.
Done when
Step green; pytest + ruff green.
Blocked by
#74.
Activity
Simon-AI-coding commentedon Aug 26, 2026
Advisor note from #74 (2026-08-26)
Decided while #74 was locking the Core
Sessionrow:name: str | None— the agent's own name, filled fromSessionInspection.name(Claude: rostername; Codex: daemonThread.namewhen known). It is an observed fact, so Build 1/7 — Discovery, roster and reachability: lock the one Session-inspection type #74 keeps it rather than dropping it at the Core boundary.label: SessionLabel | Nonestays optional and Build 1/7 — Discovery, roster and reachability: lock the one Session-inspection type #74 does not fill it — it is this ticket's composed<project> · <title>value.match_label()and the Stop Notice text fall backlabel → name → targetfrom Build 1/7 — Discovery, roster and reachability: lock the one Session-inspection type #74 on, so a discovered Session is voice-addressable by its official name before this ticket lands. This ticket's job is then to stabilise the name and compose the project half — not to plumb a new field through three layers.CONTEXT.mdlists label under Session Name's Avoid. When this ticket touchesSessionLabel/match_label/router.py/bridge.py/persistence.py, bring the vocabulary in line with the glossary in the same change, once — do not leavenameandlabelas two long-lived fields meaning almost the same thing.Simon-AI-coding commentedon Aug 26, 2026
Advisor note (2026-08-26) — an exit item inherited from #76
#76 exited with claude
progressgreen and codexprogressgated here: the acceptance drives no codex turn untilstable namepasses (journey.py:433-441), and an unnamed Codex thread reportsname: nulluntil this ticket's self-report rule exists. Done-when addition: in this ticket's acceptance run, codexprogressis green (run20260826T070925Zon #76 is the before-evidence). #76 already delivers the daemon join, the thread identity, the reader and the bound; only the name is missing.Watch for, unconfirmed: on that run
thread/read includeTurns:truefailed 3/3 on the daemon-held live thread (list_turns is not supported yet) and succeeded on the same thread once the TUI had exited. If this ticket's run reproduces it, file it then with the mechanism; it would mean the deep read fails exactly while a Session works.Simon-AI-coding commentedon Aug 26, 2026
Correction (advisor, 2026-08-26): the before-evidence run for the codex
progressexit item is20260826T074208Z(bundle byte-identical to2dfcefe, the tree PR #101 merges), not20260826T070925Z. The two runs are identical step-for-step on both lanes.Ask one Session how far it has got, and bound the answer once for bot…
Carry the user's words over a Session's own inbox, and claim only wha…
Simon-AI-coding commentedon Aug 26, 2026
Advisor ruling (2026-08-26) — one name field, and no unnamed Sessions outside a repo
The
label/namerename collapses to one field:SessionLabel→SessionName,SessionInspection.name: SessionName | Nonecomposed by the lane,core.Session.namethe only field (labeldropped),match_label→match_name, wire"name"carries<project> · <title>, fallback chain name → target. Freezing: first non-Nonewins per exactSessionTarget(legacystore.py:1875-1902); a target change (Codex pid-only row gaining its thread id;/newmoving a pid to a new thread, observed on #77) re-composes.Project half: git common-dir name when the workspace is a repo, else the workspace directory's basename;
name=Noneonly when there is no workspace. Adapted from legacylabels.py:28-70, which raised on a non-repo — v1.0 covers every Session the user starts, so a non-repo Session must stay voice-addressable. Same-basename collisions are the existingAmbiguousNameError. The git call lives once inadapters/agent/_project.py, cached per realpath'd workspace;_naming.composestays pure. Unnamed Codex rows take<project> · <first 8 of thread id>once an id exists (P10); children are unnamed. Catalogue ids renamed with the prose, old→new listed in the resolution.Simon-AI-coding commentedon Aug 26, 2026
Advisor (2026-08-27):
_naming.composereturnsSessionName | None— the body's raising signature is amended. A title that fails legacy's validation (empty, multi-line, carries the separator) yields an unnamed row, logged once, never an exception: the callers are per-row passes on the whole-machine cadence, and one bad string must not cost an enumeration. Adapted fromlegacy@1d32845:bridge/labels.py:100-105, which raised because its only caller was the self-report ingress. Reviewer finding closed by ruling. Also landed: a Child Process is never named (Core-side, holds regardless of how #79 finds children).Simon-AI-coding commentedon Aug 26, 2026
Advisor (2026-08-27) — exit items answered; two reds ruled outside this ticket
Run
20260826T122216Z(bundle =76b9c98): codexprogressgreen — #76's inherited exit item is discharged. Codexstable namegreen as'workspace-codex · 01a03e08': the shared daemon held the thread ~2 s after the TUI started, before any turn, so the<project> · <first 8 of thread id>rule applied with nothing asked of the Session. Claudestable namenow carries the project half.list_turnswatch item — reproduced once, mechanism corrected, no issue.thread/read includeTurns:true(andthread/resume) answeredlist_turns is not supported yetat 12:25:59, when the thread had zero turns; after the first turn ended (12:26:17) the deep read succeeded with the TUI still live. So it is a first-read condition on a turnless thread, not "fails while a Session works" as hypothesised on #76/#82. The lane already answersNone, logs, and keeps the cheap projection. Closed here; #82 cross-noted.Ruled outside #78: codex
approvalred — Codex raised no permission because the journey's instruction (a Write inside the workspace) is one Codex's default policy does not ask about; acceptance-design gap, filed as a ticket and sequenced before #79, since its green is the proof of #77's Codex approval route.childred both lanes — #79. Probe 0b red (0 remote audio frames in 30 s) — engine-free, unattributed, reported to Simon.Simon-AI-coding commentedon Aug 26, 2026
Resolution — PR #106, branch
78-session-namingProvenance. Base
e2090aa. Standards review23089836-b11a-4163-a42a-2321b66cb897(next: fix and stop, all fixed). Spec reviewc10107ad-b977-4933-8110-b4b76236bd0c— 3 findings, re-reviewed 2 CLOSED / 1 RETAINED, the retained one closed by the advisor ruling below. Acceptance run20260826T122216Z, bundle byte-identical to76b9c98, claude 2.1.246 / codex-cli 0.149.1.Rulings cited in the code. The
label/namecollapse into one Session Name, and the non-repository fallback to the workspace basename — adapted, because legacy raised there and v1.0 covers every Session the user starts (2026-08-26).composeansweringNonerather than raising — the ticket's-> SessionLabelamended, because legacy's caller was one self-report ingress and this one is a per-row pass over the whole machine, where one bad string must not cost an enumeration (2026-08-27). Both are quoted at the code that implements them.Shape as built.
_naming.compose(project_name, task) -> SessionName | None, pure, 64 lines._project.ProjectNamesis the one place either lane runsgit, cached per realpath'd workspace.SessionInspection.namecarries the composed value;SessionRegistryfreezes the first one per exactSessionTarget, re-composing only when the identity itself changes (#73's thread id arriving, #77's/new). Children are never named — the rule sits in the registry, so it holds however #79 finds them. Left behind as the ticket required:ai-title, mutable history, durable label rows, and the legacy self-report ingress.Exit item 1 —
stable name, red before and green after, both lanes20260826T074208Z20260826T122216Z'workspace-claude-25'— no project half'workspace-claude · workspace-claude-5b', 3 reads and a Stop'workspace-codex · 01a03e08', 3 reads and a StopExit item 2 — codex
progressgreenRed in the before-run (blocked behind the missing name, which stopped the lane at
stop noticeand skipped the rest). Green here. The codex lane went from 1 PASS / 3 FAIL / 5 SKIPPED to 7 PASS / 2 FAIL; claude is 8 PASS / 1 FAIL.Still red, neither this ticket's:
childon both lanes (#79, unbuilt) andapprovalon codex (#105 — codex raised no permission at all for the relayed instruction;approval.nonetwice in the journal).Exit item 3 — the
list_turnswatch item: reproduced once, not in the shape fearedFrom
engine-codex/engine.logandjournal.jsonlof20260826T122216Z:roster: the harness's oracle has no session id (no rollout — no turn yet), while the engine's roster already names itcodex:01a03e08-…:78415. The shared daemon held the thread ~2 s after launch, with an id andname: null.thread/read includeTurns:true→list_turns is not supported yet;thread/resume→ the same message, the same secondstable name: three reads and a Stop, all'workspace-codex · 01a03e08'progressgreen off a successful deep read — with the TUI still liveTwo log lines for the whole run, both at 12:25:59. So it is not "the deep read fails exactly while a Session works": it succeeded with the TUI live and up. The one fact true at 12:25:59 and false afterwards is that the thread had zero turns.
read_threadalready answersNoneand logs, and the roster row keeps its cheap projection, so there is no user-visible cost. Ruled no issue (advisor, 2026-08-27) and corrected on #82.That same timeline is also why the codex lane could be named before its first turn: the daemon had the thread id when nothing else did, which is exactly what
<project> · <first 8 of the thread id>is for.Gates.
ruff check,ruff format --check,pytestgreen locally; CI on #106.Give a Session one name, composed once and never changed again (#78) (#…
Drop the threads codex runs for itself before they become Sessions (#112
Rebase onto #110, bring its sixth chat read under the rule, and corre…
Attribute a chat message to the Session it names (#109) (#115)
Refuse the prompt codex hands back as a name, and follow the rename (#…
The task half of a Session Name describes the Session and climbs (#273,