Skip to content

/office-hours checks for a real user first and ends startup sessions with a dated test (GST-3024) - #4468

Open
garrytan wants to merge 17 commits into
mainfrom
office-hours-partner-sessions
Open

garrytan wants to merge 17 commits into
mainfrom
office-hours-partner-sessions

Conversation

@garrytan

@garrytan garrytan commented Nov 17, 2026 •

Copy link
Copy Markdown
Owner

/office-hours now opens every startup session with one reality-check question (a paying user, or someone who used it this week unasked) and sends founders with no outcome yet, or a new Pivot goal, to a one-page decision memo instead of a design doc. Every startup session ends with a dated KILL/CONTINUE or SUCCESS line, which the next session on that project brings back once the date passes.
Why: founders were getting a polished design doc before anyone had used or paid for what they were building.
Nothing to do. #4463, which this builds on, is already on main.

What changed

Founders kept getting a polished design doc before anyone had paid for or used what they were building. Now the first question asks for one real outcome. Without one, the session becomes a memo session. It brings a ceiling check for pivots, a conversation quota with the math, a payment test, and a memo with one action and a kill date. Memos are saved as *-memo-*.md (repo copy in docs/memos/), so the design-doc lookup and plan reviews never read one as a design, including on branches whose names contain "design" or "memo".

The approved KILL/CONTINUE or SUCCESS line is recorded, and gstack-developer-profile --kill-check <slug> brings it back at the start of the next startup session once its date has passed (local date; an empty slug is an error instead of "nothing due"). Customer names and prices stay in the private ~/.gstack copy; repo copies and second-opinion summaries use roles or initials. Builder mode is unchanged. The competitor search moved into its own section, and a "pivot" goal joins the question registry and signals.

This head rewords a set of coaching lines in plainer, general terms and shortens ten of the new lines so the merged text fits the size cap. Each technique keeps its meaning. That covers the ceiling check, the evidence ladder, the conversation quota, the payment test, the smallest first version, delivering by hand first, the pivot questions, memo routing and the reality check's skip and weak-answer rules.

Blast radius

The prose changes stay inside /office-hours. On the merged tree its text is 159,201 bytes against an unchanged cap of 159,285 (131,110 × 1.2149), and the always-loaded part is 44,256 bytes, where all growth since the last review came from main itself; the skill-size, parity, context-budget and golden guards passed at this head. gstack-design-doc-find has three callers (discovery, /autoplan, plan-design-review), all reading one path from stdout; repo docs/designs lookups are unchanged and only private state-root names skip memos (tests ran, and a smoke probe confirmed it).

Test Coverage

Coverage audit (audit only, no tests written): 9 of 14 changed code paths have a test (64%), 8 of them ★★★. Tests: 1837 → 1837 (+0 new). Gaps left as proposed tests: a commitment dated today reads as due, a commitment row missing its test or date is ignored, reconcile ignores commitment rows, and a registry check for the new pivot option. The prompt behavior (reality check, memo and pivot routing, kill-date greeting) has no free test; it is a paid-eval case.
Tests ran: free 1718 (files, CI run 44649776311)

Pre-Landing Review

One read-only full review of this head: testing, security and state-machine specialists, a red team, a Claude adversarial pass and a second-model review (GPT-6.1 Sol), which came back clean. No finding blocks: none would fail a test, break production or lose data. Nothing was fixed in this pass, by instruction; the advisory items are listed below. A separate read of every reworded and shortened line found no technique that lost its meaning.

Review notes

Below the blocking floor (advisory, informational, test-gap or wording items); recorded here and not fixed in this PR.

  • INFORMATIONAL sf-ad11d12eb79b security finding, details recorded privately
  • INFORMATIONAL two red-team findings, details recorded privately
  • INFORMATIONAL bin/gstack-developer-profile:507 A newline in kill_test forges KEY: VALUE output lines (KILL_DUE: no before the real KILL_DUE: yes).
  • INFORMATIONAL bin/gstack-developer-profile:508 A non-ISO kill_date never comes due; nothing validates the field.
  • INFORMATIONAL bin/gstack-developer-profile:508 KILL_DUE turns yes on the kill date itself while the header, changelog and docs/skills.md say once the date has passed.
  • INFORMATIONAL office-hours/sections/design-and-handoff.md.tmpl:234 kill_test may hold customer names into developer-profile.json, which --log-session queues for brain sync; same path assignment text already uses.
Documentation

Scope: this branch against 8586794: 22 changed files. Docs that users and contributors read: docs/skills.md, README.md, the changelog fragment, TODOS.md, plus the office-hours authored templates.

Result: current. docs/skills.md already explains the startup reality check, the Pivot goal, the decision memo with its *-memo-*.md naming, dated assignments, and gstack-developer-profile --kill-check. The changelog fragment matches what the change does. The README needed no change. The audit was read-only, so no files were edited.

Debt: one sentence in docs/skills.md, "Both modes end with a design doc", is now true only with the qualifier in the next paragraph. Rewording it is optional.

Tests
  • Free Tests run 44649776311 at 8fcca8a: 1718 files passed, reused by the final check
  • Windows Smoke run 44649776178 and macOS named regressions (run 44649776311): passed
  • Local guards at 8fcca8a: typecheck, skill-size, parity-suite, context-budget, module-size, goldens and conflict markers passed
  • Coverage audit at 8fcca8a: design-doc-find and developer-profile 104 pass, 0 fail; skill-validation (office-hours) 53 pass, 0 fail
Why

"Founders were getting a polished design doc before anyone had used or paid for what they were building." Now a session with no real outcome ends in one action with a dated test instead.

Exploratory QA

CLI smoke on the changed commands in an isolated state root: 4 probes, 4 pass. A commitment dated today reads KILL_DUE: yes; a commitment row does not count as a session or move the last assignment; a decision memo is never returned as a design doc. Not probed: a non-ISO kill date and a multi-line test. The prompt behavior has no CLI probe.

Eval Results

Prompt files changed, but no paid eval ran at this head, by instruction. The PR eval run at 07dc0c5 (run 44192910040) passed 3 of 3 selected cases.

Review bot comments

Review bot triage: not run in this pass.

Scope Drift

Scope Check: CLEAN. The three commits since the last full review reword coaching lines, merge main, and shorten lines to fit the size cap.

Verification Results

4 PASS, 0 FAIL, 0 BLOCKED (the smoke probes above). No plan checks were owed.

Issues

Refs #4463

TODOS

Removes the TODOS.md item for setting the /office-hours size ratio to its measured value; this PR sets it.

Agent notes

Agent notes

Plan: approved: run-yibgts-20261117204732 (Garry "do it and do it", 2026-11-17). Deviations from plan-v2 as approved: kill-date follow-up via a separate --kill-check; wallet and two-sentence are pushes inside Q4 and Q1; memo paid coverage is the sealed gstack-evals eval. maxSizeRatio 1.1993 -> 1.2149 approved by Garry (U1, 2026-11-17, via GSTA-36).

Approvals

Approved: yes (record r-ba1f5129bacc). The full approval record stays private.
Plan: approved (record r-d3a4125ef6f6)
Plan check: approved (record r-d3a4125ef6f6)

REVIEW_LOOP: passes=1 blocking_found=0 deferred=0 noted=5 fixed=0 scope=frozen ledgers_read=3
Gate edits: none detected (7 listed, 4 read; patterns: RH-1,RH-3,RH-4,RH-12,RH-13,RH-14,RH-15,RH-16; unlisted idioms: none)
Base merge (rule-19 grant): Captain GStack Minder (GSTA-36), 2026-11-19: "merge origin/main into #4468 so the size cap is measured and reviewed on the tree that lands; main's own office-hours text grew 394 bytes since fc5c285." Merged in 767bb1d.
Review rule: Captain ruling under R51: "one read-only full review per R51, with no fix round and any blocker coming to me". Ran once at 8fcca8a; no blocker; no fix commits.
Meaning check: every reworded (3966e66) and shortened (8fcca8a) line read against its earlier text; no technique lost its meaning. Pattern 9 no longer mentions the investor question, and the memo route is now pointed to as "below"; both keep the technique.
Issue links: hand-set to Refs #4463. gstack-issue-links also listed Fixes #4424 and about 60 Refs from pre-squash commits of #4463's branch that this branch carries; none belong to this PR.
Paid eval: none at this head (no paid evals or screens, per the brief). Earlier: PR eval lane run 44192910040 at 07dc0c5, 3 of 3 passed.
Pictures: none added for this head.
Refs GST-3024. Lane GSTA-595.

{
  "schema_version": 1,
  "pr": 4468,
  "head": "8fcca8a5a0491ca6077b28103d85a47938fa6264",
  "version": "1.102.0.0",
  "base": "main",
  "gate": {
    "free": {
      "pass": 1718,
      "fail": 0,
      "skip": 0
    }
  },
  "ci_carried": [
    {
      "lane": "free",
      "run": 44649776311,
      "attempt": 1,
      "tree": "23e98c052be5a96c35f98e774a00c29da265a280"
    }
  ],
  "calls": 19,
  "session_kind": "unattended",
  "run": "/home/user/.capy/work/gstack/ship-4468-8fcca8a5",
  "head_tree": "23e98c052be5a96c35f98e774a00c29da265a280",
  "product_patch_id": "48b0f022baed0c865496b6ec389b9006e6ecece2",
  "gstack_version": "1.102.0.0",
  "review_depth": {
    "depth": "full",
    "class": "gate",
    "runtime_lines": 258,
    "rules": "3e7e9d5d0a0d8879",
    "base": "858679423cc4778f9961d4aaf3c061a9368985ec",
    "registry": "858679423cc4778f9961d4aaf3c061a9368985ec",
    "gate_scan": {
      "tagged": 0,
      "head": "8fcca8a5a0491ca6077b28103d85a47938fa6264"
    },
    "reasons": [
      "gate-class",
      "runtime=258"
    ],
    "risk": [],
    "mode": "panel"
  },
  "pregate": "requires-remote cleared @23e98c052be5 (8 checks, current for 8fcca8a5a049)",
  "requires_remote": [
    "free-tests.yml/macos-named-regressions: https://github.com/garrytan/gstack/actions/runs/44649776311/attempts/1",
    "windows-smoke.yml/windows-smoke: https://github.com/garrytan/gstack/actions/runs/44649776178/attempts/1",
    "pregate.json/remote_gate: remote-gate tree=23e98c052be5 tests=38501 skipped=104 runs=44649776311/1 profiles=ci-reuse/free-tests heads=8fcca8a5a049 verified=2026-11-20T01:17:01.728Z"
  ],
  "issues": {
    "fixes": [],
    "refs": [
      4463
    ],
    "flakes": []
  },
  "docs": {
    "status": "current",
    "scope": "docs/, README.md, AGENTS.md, CLAUDE.md, changelog.d fragment and the office-hours templates"
  },
  "review_loop": {
    "passes": 1,
    "blocking_found": 0,
    "deferred": 0,
    "noted": 5,
    "fixed": 0,
    "scope": "frozen",
    "rounds": 1,
    "ledgers_read": 3
  },
  "gate_dispositions": []
}
{"head":"8fcca8a5a0491ca6077b28103d85a47938fa6264","run":"ship-4468-8fcca8a5","status":"started","review_loop":{"passes":1,"blocking_found":0,"deferred":0,"noted":1,"fixed":0,"scope":"frozen","rounds":1}}
{"head":"4afa7f2b3e291758f6a7842dd066cde22606fca1","version":"1.100.0.0","gstack_version":"1.100.0.0","run":"ship-4468-0787a1e","review_loop":{"passes":1,"blocking_found":2,"deferred":0,"noted":2,"fixed":6,"scope":"frozen","fix_checks":2,"rounds":3,"ledgers_read":6,"dispatches":{"specialist":3,"red-team":1}},"depth":"full"}

garrytan and others added 11 commits November 17, 2026 21:54
…ad of the prompt (GST-3024)

The 34-entry resource pool moves from office-hours/sections/design-and-handoff.md into lib/founder-resources.ts, served by bin/gstack-founder-resources (index lists unshown entries from the builder profile, show prints the chosen ones). Behavior is unchanged; the skill's union shrinks by 8,504 bytes.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
…did (GST-3024)

The command now prints every entry not yet shown, blurbs and URLs included, so picking resources sees what it saw before; it drops the index/show split, reports PROFILE_READ: failed when the builder profile can't be read, and its argument handling is tested through runScriptSync.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
…(GST-3024)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
…urce move (GST-3024)

The previous commit picked up four section files that belong to the follow-up content change. Nothing loads them yet; they return with that change.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
… the Pivot goal signal (GST-3024)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
…ntent (GST-3024)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
…sessions

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

# Conflicts:
#	test/founder-resources-bin.test.ts
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
@capy-ai
capy-ai Bot changed the base branch from office-hours-resources-lookup to main November 17, 2026 22:49
@trunk-io

trunk-io Bot commented Nov 17, 2026

Copy link
Copy Markdown

Merging to main in this repository is managed by Trunk.

  • To merge this pull request, check the box to the left or comment /trunk merge below.

After your PR is submitted to the merge queue, this comment will be automatically updated with its status. If the PR fails, failure details will also be posted here

garrytan and others added 3 commits November 17, 2026 23:42
… design-doc lookups (GST-3024)

The reality check now derives the project slug itself, --kill-check refuses an empty slug and compares dates in local time, the commitment log line says how to quote the founder's own words, and gstack-design-doc-find skips memos on branches whose names contain 'design'.

Review-Finding: q1/specialist-batched#1 reader=security
Review-Finding: q1/specialist-batched#2 reader=testing
Review-Finding: q1/red-team#1 reader=red-team
Review-Finding: q1/red-team#2 reader=red-team
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
…GST-3024)

The memo filter in gstack-design-doc-find now checks only the file name of private state-root docs and no longer touches docs/designs, so a checkout or topic slug containing '-memo-' keeps its design doc.

Review-Finding: q1v/adversarial-native#1 reader=adversarial-native
Review-Finding: q1v/outside-voice#1 reader=outside-voice
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
… the skills guide (GST-3024)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
capy-ai Bot added a commit to garrytan/gstack-evals that referenced this pull request Nov 18, 2026
…e during authoring, so the content change gets no verdict (#368)

Adds the results of the office-hours rewrite eval. Authoring stopped at its cap with 19 of 110 sealed briefs placed, so no session ran and gstack's office-hours content change (garrytan/gstack#4468, head 4afa7f2b) gets no verdict from this eval.
It also lands the eval's harness and the five dated plan changes the paid run used, and it prices two possible reruns for whoever decides on one.
Please merge with a merge commit, not a squash: the paid run checked out c008b08 on this branch, so that commit has to stay on main.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
@capy-ai

capy-ai Bot commented Nov 19, 2026

Copy link
Copy Markdown
Contributor

Held on purpose: this rewrite of /office-hours merges only after gstack-evals' redesigned /office-hours eval reports, because the maintainers want measured evidence that the rewrite helps before it ships. Green checks alone won't merge it.

garrytan and others added 3 commits November 20, 2026 00:47
…on quotes and numbers (GST-3024)

The pushback patterns, the pivot flow and the discovery flow keep every technique but no longer carry quoted phrasing or figures from the source material; the size-cap comment drops its source label.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
Captain GStack Minder (GSTA-36), 2026-11-19: merge origin/main into #4468 so the size cap is measured and reviewed on the tree that lands; main's own office-hours text grew 394 bytes since fc5c285.
…cap after main grew (GST-3024)

Main's own /office-hours text grew 394 bytes since this branch forked, so ten of this change's added lines are shortened by 473 bytes with the same steps and meaning; the cap stays at 1.2149.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

/ship test step fails on Linux when the unit suite has thousands of files (argument list too long)

1 participant