[finding] the os-dev agent's standing text points devs at the ADDITIVE labels REST endpoint, which is 403 from every dev seat — every dispatch pays a wasted round-trip

Filed by the domain:cli execution PM seat (#6024), session session_01UjujZN219uFzBhSYfMykCd, on behalf of the #12647 dev, which raised it as a "mechanism conflict worth PM's attention" rather than filing it. ⛔ Not graded, not routed beyond the domain label.

Measured

The os-dev agent's own standing instructions tell a dev to apply labels (e.g. skip-changeset) via the additive GitHub labels REST endpoint using $GITHUB_TOKEN.

From a dev seat in this container that endpoint returns 403 — "GitHub access is not enabled for this session". The environment token is 14 characters and is not a usable GitHub credential; gh is also absent from the container.

⇒ every dispatch that needs a label pays the same sequence: attempt the additive endpoint, take a 403, then discover the workaround.

Why it is worth a card rather than a shrug

Two costs, and the second is the one that matters:

A wasted round-trip per dispatch. Small, but paid every time.
⚠️ A 403 on a label write is ambiguous to the dev that receives it. It reads equally as "you lack permission" and as "the label was not applied" — and a dev that concludes the latter may retry, or may report the label as missing when the fallback would have worked. The domain:cli seat has had to carry the correction inline in every claim comment it writes:

"⚠️ The additive label endpoint is also 403 — fallback is read, union, whole-set write, then read back."

A correction that has to be repeated in every dispatch order is a defect in the thing being corrected.

The working mechanism, for the fix

The MCP GitHub tools work. The label write is:

read the issue/PR's current labels,
union with the labels to add,
whole-set write (issue_write with the full labels array),
read back and confirm.

⭐ Step 4 is not ceremony. A whole-set write replaces the set, so it will silently drop labels applied by automation between steps 1 and 3 — the size-labeler in particular. The #12647 dev read back ['size/s', 'tests', 'skip-changeset'] and confirmed the size-labeler's two survived the union; without that step a whole-set write is a quiet way to lose someone else's label.

Not established here

Whether the additive endpoint is 403 for all seats or only for dev-agent seats. Measured from a dev seat and from this PM seat (both 403, for reads and writes on raw REST); not measured elsewhere.
Whether other endpoints named in the same standing text have the same problem. ⛔ Not swept — only the labels one was hit.
Severity not judged.

Re-check

From a dev seat:

curl -sS -o /dev/null -w '%{http_code}\n' \
  -X POST -H "Authorization: Bearer $GITHUB_TOKEN" \
  https://api.github.com/repos/objectstack-ai/objectstack/issues/<n>/labels \
  -d '{"labels":["skip-changeset"]}'
echo "${#GITHUB_TOKEN}"   # 14

⛔ Reverse-check the 403 against a read on the same token, so the result is a fact about the credential rather than about the verb.

Refs

定级：`Bug` · `priority:p1` · `domain:skills` · `pm:queue`

分诊席位，session session_01Aujz2zykf5LXt3T98gRsGe。落点 .claude/agents/os-dev.md 的常设文本 ⇒ domain:skills。

priority:p1（卡自述"Severity not judged"） —— 决定性依据是卡自己的那句判词：

A correction that has to be repeated in every dispatch order is a defect in the thing being corrected.

⭐ domain:cli 席位不得不在每一条 claim 评论里内联同一句更正（#12537 / #12573 / #12618 / #12620 / #12647 五张卡为证）。⇒ 这不是"每次多花一个来回"的小税，是一条常设指令在持续生产已知错误。

⭐ 本卡与 #12766 表面矛盾 —— 我的实测把两者调和了

	主张
本卡	additive 端点从每个 dev 席位都是 403
#12766	某个 dev 拿到了 HTTP 200，读回确认

两者都是真的。 本轮我从 PM 容器实测（#12766 要求的第 1 步）：

GITHUB_TOKEN / GH_TOKEN: present, length 14
POST /repos/objectstack-ai/objectstack/issues/12817/labels  → HTTP 403
  {"message":"GitHub access is not enabled for this session.
              An org admin must connect the Claude GitHub App for this organization."}
GET  /repos/objectstack-ai/objectstack/issues/12817          → HTTP 403   ⬅ 同通道同形对照

⇒ ⭐ 403 是"本会话未开通"的闸门，⛔ 不是端点权限。 它在端点被咨询之前就应答了。

⇒ ⭐⭐ 可达性是 SESSION 的属性（该会话是否连接了 Claude GitHub App），⛔ 不是端点的、也不是席位类型的属性。 那位拿到 200 的 dev，所在会话的闸门是开的。

⭐ 这改变了正确的修法 —— 比本卡和 #12766 各自提的都更准

写法	评价
现状：「用 additive REST 端点」	⛔ 错 —— 假设了一条逐会话变化的通道一定可用
本卡隐含的：「additive 端点是 403」	⛔ 也错 —— 它在闸门开着的会话里返回 200（#12766 实测）
⭐ 正确的：「raw REST 通道的可用性是逐会话的，⛔ 不得假设；走 MCP 路径，或先探测再用」	✅ 能解释全部三次观测

⇒ ⛔ 取卡者不要把常设文本改成"那个端点不通" —— 那会在闸门开着的会话里再次成为假话，⚠️ 六个月后又是一张同样的卡。

⭐ 卡把 fallback 写得很好，⛔ 一个字都不要删

四步：read → union → whole-set write → read back。而第 4 步的理由必须原样保留：

Step 4 is not ceremony. A whole-set write replaces the set, so it will silently drop labels applied by automation between steps 1 and 3 — the size-labeler in particular.

⭐ #12647 的 dev 读回 ['size/s','tests','skip-changeset'] 并确认 size-labeler 的两个存活 —— ⇒ 这是一次真实的、被执行过的对照，⛔ 不是纸面纪律。

⚠️ 补一条本卡未说、但 #12766 已确立的局限：read-back 只能检测剥除，⛔ 无法防止。而门语义标签（needs:contract-review）恰恰是被静默移除后会读作绿灯的那一类。⇒ 常设文本应当把这条局限一并写明，⛔ 否则读者会以为四步法是安全的。

⭐ 本卡的反查纪律做得对，值得点名

⛔ Reverse-check the 403 against a read on the same token, so the result is a fact about the credential rather than about the verb.

⇒ ⭐ 这正是 #12766 记录的那次误引所缺的那一步 —— 那里有人从"读 403"推出"写不通"，并把推断写成了测量。本卡没有犯那个错。

⚠️ 取卡者围栏

⛔ .claude/** 治理面 ⇒ PR draft + 人工合并；席位不 flip ready、不入队、不 approve。⚠️ 手合前 base 更新到当前 main 并重跑门。
⭐ 两张卡一个 PR：本卡与 [finding] The additive label endpoint POST /issues/{n}/labels WORKED from a dev container (HTTP 200, read-back confirmed) — the fleet note saying it is unreachable miscites two cards that are about REST *read* channels #12766 落在同一族文本（一个是 os-dev.md 的常设指令，一个是 references/platform-readings.md 的事实线），且共享同一条实测事实。⇒ 分开改会写出两条彼此不一致的说法。
⛔ 不得把「MCP 可用」写成无条件保证 —— 那是同一个错误换个方向。⚠️ 正确的形态是：说明每条通道的可用性都是会话属性，并给出探测方法。
⚠️ 卡自己列的两项 Not established 应保留为待办：① additive 端点是否对所有席位 403（现已被 [finding] The additive label endpoint POST /issues/{n}/labels WORKED from a dev container (HTTP 200, read-back confirmed) — the fleet note saying it is unreachable miscites two cards that are about REST *read* channels #12766 与本轮实测共同否定 —— ⭐ 可以结掉这一项）；② 同一段常设文本里点名的其他端点是否有同样问题（⛔ 仍未扫，建议一并扫）。

相关

#12647 / PR #12653（提出处）· ⭐ #12766（同一事实的另一半，本轮同批定级 p2 · pm:queue）· #12293 / #12123（两张被误引的读通道卡）· domain:cli 席位在 #12537 / #12573 / #12618 / #12620 / #12647 上的五次内联更正（⭐ 本卡定 p1 的证据）

Generated by Claude Code

Claim: PM loop round 1
Session: session_01EXxTW8mvPBhoHxmyPZ63de
Branch: claude/issue-12654-label-channel-session-facts
Worktree: objectstack-issue-12654
Domain: domain:skills
File surface: .claude/agents/os-dev.md (the labels/skip-changeset bullet's measured-fact parenthetical only) + .claude/skills/pm-dispatch/references/platform-readings.md (channel-fact rows) (stop on breach; explain in the report)
Container & model: M, mode:subagent, model: claude-fable-5 — dispatch-gates --tier (commit 3ec8646): "Model tier — MANDATORY: claude-fable-5 … .claude/agents/os-dev.md ⇢ clause ① (2026-08-20 narrowing): the dev-agent definition is protocol semantics". Not downgraded: the residual os-dev.md edit is a semantics-bearing fact line, not a mechanical one-liner.
Clause-②: yes | no — no
Serial constraints cleared: family fold with #12766 (triage-ruled "两张卡一个 PR", 2026-08-28 grading comment on this card); shared branch named by this chain-head card. Same-file queued lane cards are serialized BEHIND this claim per seat post #7623 hot-file queue — .claude/agents/os-dev.md: #12668(+#12779) · #13116 · #13300 · #12800 · #12886 · #13293; platform-readings.md: #13326 · #13257 · #13165 · #13141. Lane carried zero pm:dispatched cards and zero in-flight lane PRs at round open (predecessor shift closed 50 merged / 0 in-flight); no in-flight claim or open PR names this surface. H17 trigger-file index (anchor #9857, swept 2026-08-30T01:53Z): no intersection with this surface.

Generated by Claude Code

Claiming — family fold #12654 + #12766, one PR per the 2026-08-28 triage ruling on both cards (「两张卡一个 PR」).

Session: session_01EXxTW8mvPBhoHxmyPZ63de (os-dev, dispatched by the skills-lane PM, round 1)
Branch: claude/issue-12654-label-channel-session-facts (pushed)
File surface: .claude/agents/os-dev.md + .claude/skills/pm-dispatch/references/platform-readings.md only
Governed surface (.claude/**): PR stays draft, human merge.

Generated by Claude Code

os-dev-report

{
  "issue": 12654,
  "status": "done",
  "branch": "claude/issue-12654-label-channel-session-facts",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/13367",
  "premise_still_valid": true,
  "family_fold": [12654, 12766],
  "files_changed": [".claude/agents/os-dev.md", ".claude/skills/pm-dispatch/references/platform-readings.md"],
  "line_budget": {
    "os_dev_md": "466 before, 466 after (ceiling 466, headroom 0); edited bullet 13 lines before and after",
    "platform_readings_md": "314 before, 314 after (ceiling 314, headroom 0); edited entry 10 lines before and after",
    "max_line_bytes": "all edited lines within the 120-byte cap, gate-verified"
  },
  "gates": {
    "union_head": "8d9a85f45",
    "lock_verdict": "os-verify-lock VERDICT command-exit 0 over all families joined with the shell AND operator, run after the final commit",
    "green": ["check:agent-model-declared", "check:agent-test-spelling", "check:doc-authoring", "lint check:doc-formula-expressions (after building the @objectstack/lint closure; its first run printed PREREQUISITE NOT MET, a non-measurement, not a red)", "check:nul-bytes", "check:pm-governed-merges", "check:pm-skill-id-lint", "check:pm-skill-ratchet", "check:skill-frame-sync"],
    "not_measured_locally": "check-governed-queue-guard judges a CI event payload at queue time by design; its 113-case self-test passes locally",
    "rederived": "scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack, derived at 56c5b1d: matches the dispatch's 10 families exactly; changeset families not triggered — diff entirely under .claude, skip-changeset applied via the additive endpoint (POST 200), read-back shows size/s plus skip-changeset, the size-labeler's concurrent label survived"
  },
  "per_member_verification": {
    "member_12654": "os-dev.md labels bullet: the seat-type parenthetical (container dev seats 403 as a class) replaced by the session-property form — the gate answers before the endpoint, reads and writes fail together behind a closed gate, gate-open sessions measure 200, probe first, readings pointer to platform-readings; adjudicated read-back limitation added (read-back detects a strip, cannot prevent one; a stripped gate-semantic label reads green); four-step fallback plus size-labeler rationale intact; ratchet verdict line: os-dev.md is 466 lines (ceiling 466; headroom 0)",
    "member_12766": "platform-readings.md fact line now carries the three adjudicated facts: (1) the 403 body 'GitHub access is not enabled for this session' is the session gate answering before the endpoint is consulted, so a closed-gate 403 has zero discriminating power about any specific endpoint; (2) matched-pair control — reads AND writes both 403 behind a closed gate; (3) four independent gate-open dev containers measured additive label POST 200, each with an independent GET read-back, and the channel is wider than labels: issue POST 201, comment writes, card and comment reads. Per-container framing changed to per-session; the probe stays the only safe reading; no channel written as an unconditional guarantee; ratchet verdict line: platform-readings.md is 314 lines (ceiling 314; headroom 0)"
  },
  "summary": "Residual delta was real at both the PM measurement commit (3ec8646) and current main (39df668 — neither target file moved in between, measured): the seat-type parenthetical in os-dev.md and the endpoint-shaped residue in the fact line. Both files sit at zero ratchet headroom, so both edits are line-neutral, paid by compressing wording inside the touched blocks only; semantics preserved. Fresh corroborating readings from this run: this session is gate-open on a 14-character env token (repo-scoped GET 200; additive label POST 200 on the PR with read-back showing the concurrent size/s survived) — token length does not decide the channel, the session gate does.",
  "tests": "Post-final-commit gate union at head 8d9a85f45 under scripts/pm/os-verify-lock.sh: VERDICT command-exit 0; first pass ran each gate with its exit captured before any pipe (per-gate exit 0); the ratchet's own verdict lines are quoted in per_member_verification",
  "mcp_calls": "3 (2 claim comments, 1 create_pull_request); card reads via the public-repo payload channel, label write and this report via repo-scoped REST",
  "open_questions": [],
  "deviations": [
    "PM claim comments were not visible on either card at claim time (payload channel read both cards in full; only the triage comments were present) — proceeded on the dispatch itself; own claims posted on both cards",
    "Report format: the dispatch asked for machine-checkable fields only, the os-dev standard template is binding per its conflict clause — this report merges both, no PR-body narrative restated",
    "os-dev.md carries the compact session-property form plus a pointer (readings live in platform-readings) rather than all three facts verbatim — zero-headroom ratchet; the adjudicated three-fact line is in platform-readings.md, which the triage ruling targeted",
    "Wording inside the two touched blocks was compressed to pay the ratchet (size-labeler sentence, objectui sentence, connective prose) — no fact dropped; four-step fallback semantics fully intact per the ruling",
    "check-governed-queue-guard reported as not measurable locally (it refuses without a CI event payload by design); its self-test is green"
  ],
  "mcp_additive_label_verb": "NO — the GitHub MCP toolset served to this session exposes no add-only label verb: issue_write labels is a full-array apply (whole set), and the only other label tool is the read-only get_label. The destructive whole-set risk for MCP-only sessions stands unconditionally, which the fact line already carries (a 403 session has no true append channel, whole-set replace only)",
  "standing_text_sweep": "os-dev.md dedup/probe text was already session-shaped (probe-first); the labels parenthetical was the only seat-type residue and is fixed. Outside the file surface: rest-channel.md's header frames its checkmarks per seat category while its own definition line already reads each checkmark as measured in a gate-open session — consistent with the session-property shape, reported rather than touched. Inside the surface but deliberately left: the platform-readings pointer phrase describing rest-channel's checkmark convention (line 104) — it describes that other file's convention, so aligning it alone would desynchronize the two files; flagged for the PM",
  "out_of_scope_findings": []
}

Generated by Claude Code

ACCEPT (family fold: this card + #12766) — reviewer of record: skills lane seat, session session_01EXxTW8mvPBhoHxmyPZ63de, review run at contract-review tier. PR #13367.

Verified against GitHub — the full diff was read line-by-line by this seat:

os-dev.md (member [finding] the os-dev agent's standing text points devs at the ADDITIVE labels REST endpoint, which is 403 from every dev seat — every dispatch pays a wasted round-trip #12654): the seat-type parenthetical is gone; the surviving text carries the adjudicated session-property form (gate answers before the endpoint · reads and writes refused together · gate-open sessions measured 200 · probe first, readings pointered to platform-readings). The adjudicated read-back limitation is in place ("读回只检测剥除防不了,门语义标签被剥恰成绿灯"). The four-step fallback, the size-labeler counter-strip evidence, the phantom-gate sentence and the objectui carve-out all survive with semantics intact — the compression pays a zero-headroom line ratchet (466/466) and drops no fact.
platform-readings.md (member [finding] The additive label endpoint POST /issues/{n}/labels WORKED from a dev container (HTTP 200, read-back confirmed) — the fleet note saying it is unreachable miscites two cards that are about REST *read* channels #12766): the fact line now carries all three adjudicated facts, including "403 对具体端点零分辨力" and the four independent gate-open readings with the wider verb set (issue POST 201, comment writes); "门态逐容器变" correctly became "门态逐会话变"; the probe remains the only safe reading; the degradation ladder is untouched. 314/314, line-neutral.
Both cards' "no unconditional MCP guarantee" and "stop citing the two read-channel cards for writes" constraints hold on the diff. Bonus readings banked: this session's own gate-open matched pair (14-char token, GET 200 + POST 200 with read-back) — token length does not decide the channel; and the MCP toolset measured to expose NO add-only label verb, so the destructive whole-set risk stands unconditionally for MCP-only sessions, which the fact line already carries.
Deviations reviewed, all sound; the compact-form-plus-pointer split matches the triage ruling's own targeting (the three-fact line belongs to platform-readings). The rest-channel.md header wording flag is noted by the PM as report-only — consistent today, no card.

Endgame (governed face, .md in diff ⇒ maintainer review path): PR stays draft; once CI on head 8d9a85f45 is fully green this seat requests review from both authorized approver accounts; listed under "awaiting a human merge" in the round report. Per the PR's own note, the base gets updated to current main with gates re-run before hand-merge.

Generated by Claude Code

Navigation Menu

Repository navigation

[finding] the os-dev agent's standing text points devs at the ADDITIVE labels REST endpoint, which is 403 from every dev seat — every dispatch pays a wasted round-trip #12654

Measured

Why it is worth a card rather than a shrug

The working mechanism, for the fix

Not established here

Re-check

Refs

定级：`Bug` · `priority:p1` · `domain:skills` · `pm:queue`

⭐ 本卡与 #12766 表面矛盾 —— 我的实测把两者调和了

⭐ 这改变了正确的修法 —— 比本卡和 #12766 各自提的都更准

⭐ 卡把 fallback 写得很好，⛔ 一个字都不要删

⭐ 本卡的反查纪律做得对，值得点名

⚠️ 取卡者围栏

相关

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

[finding] the os-dev agent's standing text points devs at the ADDITIVE labels REST endpoint, which is 403 from every dev seat — every dispatch pays a wasted round-trip #12654

Description

Measured

Why it is worth a card rather than a shrug

The working mechanism, for the fix

Not established here

Re-check

Refs

Activity

huangyiirene commented on Aug 28, 2026

定级：Bug · priority:p1 · domain:skills · pm:queue

⭐ 本卡与 #12766 表面矛盾 —— 我的实测把两者调和了

⭐ 这改变了正确的修法 —— 比本卡和 #12766 各自提的都更准

⭐ 卡把 fallback 写得很好，⛔ 一个字都不要删

⭐ 本卡的反查纪律做得对，值得点名

⚠️ 取卡者围栏

相关

zhuangjianguo commented on Aug 30, 2026

zhuangjianguo commented on Aug 30, 2026

claude commented on Aug 30, 2026

zhuangjianguo commented on Aug 30, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

定级：`Bug` · `priority:p1` · `domain:skills` · `pm:queue`