docs: re-init CLAUDE.md and align all four harness prompts - #2
Merged
Conversation
`/init` replaced the scaffold seed `CLAUDE.md` with a runtime prompt grounded in this repo, and the other three harness files were brought in line with it — each written for its own harness rather than restating the same text. - CLAUDE.md: command block (tests, single test, every lint/rubric/smoke gate), the cross-file CLI wiring (`_CliArgumentParser.error()` override, the `_json_hint` argv peek, the `parser_class` propagation trap, `CliError`, the stdout/stderr split, the catalog walked by `tests/test_cli.py`), a five-place checklist for adding a Substack noun, and `_PROMPT_FILE` vs `_RESIDENT_PROMPT`. - QWEN.md: the coding-session machinery for a Qwen Code session. - AGENTS.override.md: context only (identity stays in `.pi/SYSTEM.md`), and tells Pi's non-coding lane to report the scaffold gap rather than infer an implementation from the project description. - AGENTS.colleague.md: rewritten around colleague as the `/ask-colleague` second mind — per-verb expectations, the contracts a diff must respect, and the colleague#494 empty-skills caveat. - README.md: template-clone instructions replaced by this repo's own Status / CLI / Development sections. Fixes four drifts the docs carried: a Substack surface (posts, comments, statistics, subscribers) that does not exist on disk; a quickstart using `uv run substack-cli …` when `[project.scripts]` installs the binary as `substack`; a skill count of 11 against 19 on disk; and three files calling `CLAUDE.md` "the fullest write-up" while it was still the `/init` seed. Verified: 117 tests pass, `teken cli doctor . --strict` PASS, `harness-smoke --stage config` 6/6 PASS, scan-secrets clean, markdownlint clean, `devex pr lint` no violations. Qwen and Pi were asked "Who are you, and what can you do?" in this tree and both answered from their new files. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]> Claude-Session: https://claude.ai/code/session_01VaCYnmwSaYL2DLxDtd1fTs
Contributor
Author
|
/agentic_review |
OriNachum
had a problem deploying
to
testpypi
September 12, 2026 17:39 — with 
GitHub Actions
Failure
PR Summary by QodoAlign harness prompts with substack-cli scaffold reality
AI Description
Diagram
High-Level Assessment
Files changed (8)
|
Code Review by Qodo
1.
|
The Worktrees section claimed the vendored skill's fan-out example uses a shared `../worktrees/` path that must be overridden. It does not: the current `.claude/skills/assign-to-workforce/SKILL.md` mandates the same repo-named `.worktrees.<repo-name>` root this file does, and explicitly says "Never use a bare `../worktrees/`". The claim was inherited from culture-agent-template's CLAUDE.md and describes an older revision of the vendored file. Only the branch-name half of the override still applies (`agent/<task-id>`). Found by `ask-colleague review` on this PR. Its other two findings: the "eight skills originate in devague" phrasing is correct as written (docs/skill-sources.md lists exactly eight devague-origin rows, all re-broadcast via guildmaster — the direct-from-devague vendoring it recalled is marked superseded in that ledger), and the code-side "clonable template" strings are filed as #3 rather than folded into a doc-only PR. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]> Claude-Session: https://claude.ai/code/session_01VaCYnmwSaYL2DLxDtd1fTs
OriNachum
had a problem deploying
to
testpypi
September 12, 2026 17:44 — with
GitHub Actions
Failure
Contributor
Author
|
Ran 1. FIX — my citation of the vendored 2. Deferred to #3 — the code still self-describes as a template. 3. PUSHBACK — "eight skills originate in devague" is correct as written. Colleague read it as seven-plus-one and suggested naming the two legs. Colleague was explicit about what it did not check: it did not run the suite or Remaining CI failure is repo setup, not this branch. |
…ecision Six inline findings from the Qodo review on PR #2, each verified against disk before acting: - Mark the Substack surface `(planned)` at first mention in all four harness files and on the "Adding the Substack surface" heading, per this repo's own rule that anything ahead of disk carries the marker. - Note that `markdownlint-cli2` is an npm prerequisite, not something `uv sync` installs — CI installs it separately, so a clean checkout following the documented sequence stopped there. - Stop implying all four harnesses load the shared skill tree: the symlinks are shared wiring, but colleague 1.76.0 loads 0 of the 19 (colleague#494). The caveat was in AGENTS.colleague.md only; it is now in CLAUDE.md and QWEN.md where the "one tree, four loaders" claim was made. - Qualify the CliError contract. "Every failure raises CliError" contradicted two deliberate paths: `_CliArgumentParser.error()` emits a CliError then raises SystemExit, and `doctor` returns 1 for an unhealthy report rather than raising. Both are now named so neither gets "fixed" by a future change. - AGENTS.colleague.md claimed nothing from a read-only run reaches the asker's checkout. Tracked files, correct — but the run artifact is copied into the gitignored `.colleague/` dir for grading. Stated. - docs/skill-sources.md listed 17 rows while `.claude/skills/` holds 19: `recall` and `remember` were vendored without ledger entries. Added both (guildmaster, scope literal adapted), plus a header line accounting for all 19. Qodo read the gap as "guildmaster does not supply these two" — it does ship both; the defect was the missing rows, not the supplier. Also corrected the kit description in CLAUDE.md, QWEN.md and README.md: 19 skills is 17 guildmaster + ask-colleague direct from colleague, not a single "canonical guildmaster kit". Co-Authored-By: Claude Opus 5 (1M context) <[email protected]> Claude-Session: https://claude.ai/code/session_01VaCYnmwSaYL2DLxDtd1fTs
OriNachum
had a problem deploying
to
testpypi
September 12, 2026 17:48 — with
GitHub Actions
Failure
Contributor
Author
|
FIX — dd96960. Fair reading of this repo's own rule (anything ahead of disk carries
|
Contributor
Author
|
FIX — dd96960. Correct and worth catching:
|
Contributor
Author
|
FIX — dd96960. The caveat existed in
|
Contributor
Author
|
FIX — dd96960. Verified both paths in source:
|
Contributor
Author
|
FIX — dd96960. Confirmed empirically, not just by reading
|
Contributor
Author
|
FIX on the defect, PUSHBACK on the diagnosis — dd96960. The gap is real: But guildmaster does supply them — One thing the ledger now records that neither of us started from:
|
|
OriNachum
added a commit
that referenced
this pull request
Sep 13, 2026
…lass (v0.10.0) (#5) * docs: re-init CLAUDE.md and align all four harness prompts `/init` replaced the scaffold seed `CLAUDE.md` with a runtime prompt grounded in this repo, and the other three harness files were brought in line with it — each written for its own harness rather than restating the same text. - CLAUDE.md: command block (tests, single test, every lint/rubric/smoke gate), the cross-file CLI wiring (`_CliArgumentParser.error()` override, the `_json_hint` argv peek, the `parser_class` propagation trap, `CliError`, the stdout/stderr split, the catalog walked by `tests/test_cli.py`), a five-place checklist for adding a Substack noun, and `_PROMPT_FILE` vs `_RESIDENT_PROMPT`. - QWEN.md: the coding-session machinery for a Qwen Code session. - AGENTS.override.md: context only (identity stays in `.pi/SYSTEM.md`), and tells Pi's non-coding lane to report the scaffold gap rather than infer an implementation from the project description. - AGENTS.colleague.md: rewritten around colleague as the `/ask-colleague` second mind — per-verb expectations, the contracts a diff must respect, and the colleague#494 empty-skills caveat. - README.md: template-clone instructions replaced by this repo's own Status / CLI / Development sections. Fixes four drifts the docs carried: a Substack surface (posts, comments, statistics, subscribers) that does not exist on disk; a quickstart using `uv run substack-cli …` when `[project.scripts]` installs the binary as `substack`; a skill count of 11 against 19 on disk; and three files calling `CLAUDE.md` "the fullest write-up" while it was still the `/init` seed. Verified: 117 tests pass, `teken cli doctor . --strict` PASS, `harness-smoke --stage config` 6/6 PASS, scan-secrets clean, markdownlint clean, `devex pr lint` no violations. Qwen and Pi were asked "Who are you, and what can you do?" in this tree and both answered from their new files. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]> Claude-Session: https://claude.ai/code/session_01VaCYnmwSaYL2DLxDtd1fTs * docs: correct the assign-to-workforce worktree citation in CLAUDE.md The Worktrees section claimed the vendored skill's fan-out example uses a shared `../worktrees/` path that must be overridden. It does not: the current `.claude/skills/assign-to-workforce/SKILL.md` mandates the same repo-named `.worktrees.<repo-name>` root this file does, and explicitly says "Never use a bare `../worktrees/`". The claim was inherited from culture-agent-template's CLAUDE.md and describes an older revision of the vendored file. Only the branch-name half of the override still applies (`agent/<task-id>`). Found by `ask-colleague review` on this PR. Its other two findings: the "eight skills originate in devague" phrasing is correct as written (docs/skill-sources.md lists exactly eight devague-origin rows, all re-broadcast via guildmaster — the direct-from-devague vendoring it recalled is marked superseded in that ledger), and the code-side "clonable template" strings are filed as #3 rather than folded into a doc-only PR. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]> Claude-Session: https://claude.ai/code/session_01VaCYnmwSaYL2DLxDtd1fTs * docs: address Qodo review — planned markers, lint prereq, contract precision Six inline findings from the Qodo review on PR #2, each verified against disk before acting: - Mark the Substack surface `(planned)` at first mention in all four harness files and on the "Adding the Substack surface" heading, per this repo's own rule that anything ahead of disk carries the marker. - Note that `markdownlint-cli2` is an npm prerequisite, not something `uv sync` installs — CI installs it separately, so a clean checkout following the documented sequence stopped there. - Stop implying all four harnesses load the shared skill tree: the symlinks are shared wiring, but colleague 1.76.0 loads 0 of the 19 (colleague#494). The caveat was in AGENTS.colleague.md only; it is now in CLAUDE.md and QWEN.md where the "one tree, four loaders" claim was made. - Qualify the CliError contract. "Every failure raises CliError" contradicted two deliberate paths: `_CliArgumentParser.error()` emits a CliError then raises SystemExit, and `doctor` returns 1 for an unhealthy report rather than raising. Both are now named so neither gets "fixed" by a future change. - AGENTS.colleague.md claimed nothing from a read-only run reaches the asker's checkout. Tracked files, correct — but the run artifact is copied into the gitignored `.colleague/` dir for grading. Stated. - docs/skill-sources.md listed 17 rows while `.claude/skills/` holds 19: `recall` and `remember` were vendored without ledger entries. Added both (guildmaster, scope literal adapted), plus a header line accounting for all 19. Qodo read the gap as "guildmaster does not supply these two" — it does ship both; the defect was the missing rows, not the supplier. Also corrected the kit description in CLAUDE.md, QWEN.md and README.md: 19 skills is 17 guildmaster + ask-colleague direct from colleague, not a single "canonical guildmaster kit". Co-Authored-By: Claude Opus 5 (1M context) <[email protected]> Claude-Session: https://claude.ai/code/session_01VaCYnmwSaYL2DLxDtd1fTs * spec: substack-domain-surface (devague /scope + /think) Converged frame for the Substack control surface: post/feed/comment/ reaction/account nouns, webglass-cli as the runtime browser plane, public read verbs first, owner verbs gated on webglass authenticated sessions (agentculture/webglass-cli#17). Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * spec: challenge pass on substack-domain-surface (devague /challenge) Rigorous pass (hard-to-reverse publish + security-sensitive session): webglass has no network lens (capture via Chrome-MCP now, lens asked in webglass-cli#17), two API hosts, markdown->ProseMirror is real work, no auto-retry on writes, partial-state reporting, three-state auth probe, untrusted third-party text, containment twins for every write verb, draft-first publish with --no-email for live proofs. One self-reported lapse (l1) confirmed. Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * plan: substack-domain-surface (devague /spec-to-plan) 17 tasks in 7 waves covering all 56 spec targets: baseline, HTTP transport, webglass adapter, untrusted-text helper, Chrome capture; five nouns; post write side + markdown body; parser/learn/catalog wiring; docs, version bump, CI gates; public and owner live proofs. Owner proof (t18) is a follow-up risk on webglass-cli#17. Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * plan: implementation split (gate 2 approved) Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * t1: record pre-feature learn baseline Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * t4: untrusted third-party text renderer Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * t3: webglass subprocess adapter Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * t2: stdlib HTTP transport with host validation and GET-only backoff Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * t6: post noun read side (list, get, overview) Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * t5: account noun (whoami auth probe, overview) Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * t12: observed Substack API endpoints for the v1 verbs Captured 2026-09-13 in the owner's logged-in browser on jetsonailab.substack.com: drafts create/update/publish/schedule/ unschedule/unpublish/delete, comments create/reply/delete, post and comment reactions add/remove, public archive/post/comments reads, the subscription endpoint as the whoami source, and the substack.com reader feed. Unobserved paths are marked as such. Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * t12: reword a query-string example the secrets gate flagged Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * t5: whoami reads user_id from /api/v1/subscription Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * t9: feed noun (read, overview) Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * t7: comment noun (list, reply, delete, overview) Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * t8: reaction noun (list, add, remove, overview) Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * t10: post write side (publish, schedule, unpublish, delete) and markdown body builder Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * t13: wire the five Substack nouns into the parser, learn and explain Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * t15: bump version to 0.10.0 Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * http: descriptive User-Agent and fail-fast on non-retryable 4xx The live public-read proof returned 403: Substack rejects urllib's default Python-urllib agent (a descriptive substack-cli/<version> agent is accepted; verified with curl). The GET loop also retried every HTTPError; it now retries only 429, 5xx and transport errors, per spec claim c38. Two tests that asserted four attempts on a 404 now assert one. Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * t16: CI gates — socket-blocking fixture, repo invariants, coverage floor Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * t14: document the Substack surface across README and the four harness prompts Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * t17: public-half live proof against two publication hosts Evidence stored as a text record with a header line (JSON body): scripts/scan-secrets.py fails any JSON-parsable file whose url keys point off localhost, and the t2 invariant forbids tracked *.json files naming substack.com. Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * delivery: summarize the substack-domain-surface workforce run Partial run: 16 of 17 plan tasks delivered, t18 (owner-half live proof) blocked on agentculture/webglass-cli#17. Deviations d1-d4, lapses l1-l15, validation ledger o1-o23/e1-e22/b1-b4 all adjudicated. Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * fix: address Qodo review on PR #5 (comments 3998832718, 3998832722, 3998832725, 3998832730, 3998832732, 3998832734, 3998832736, 3998832739) - 3998832718: `comment list` now sends the observed query (all_comments=true&sort=best_first, urlencoded) and flattens nested `children` replies depth-first, parents before replies, keeping parent_id/ancestor_path on each item. - 3998832722: reaction add/remove build their URL from http.publication_base(), so SUBSTACK_API_BASE reaches writes as it does reads. - 3998832725: `feed read` builds limit/cursor with urllib.parse.urlencode, so a cursor containing &, #, + or % can no longer forge or truncate the query. - 3998832730: run_webglass passes a finite timeout (120s, SUBSTACK_WEBGLASS_TIMEOUT; invalid -> CliError(1)) and maps TimeoutExpired to CliError(2) naming the variable. - 3998832732: http passes a finite timeout to opener.open (30s, SUBSTACK_HTTP_TIMEOUT; invalid -> CliError(1)); timeouts retry on GET and raise a single CliError(2) on writes. - 3998832734: UnicodeDecodeError/JSONDecodeError around response decoding on both the GET and write paths -> CliError(2) naming the URL, never retried. - 3998832736: an argparse usage/unknown-verb answer from webglass now maps to CliError(2) saying webglass-cli has no authenticated request verb yet, citing agentculture/webglass-cli#17. The verbs stay registered. - 3998832739: --body-json must decode to {"type": "doc", "content": [...]}; anything else is CliError(1) with the ProseMirror remediation. - 3998832738 (--send): behaviour unchanged by design; the help text and docstring now say the send:true path is unverified against the live API. Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * refactor: address SonarCloud smells on the Substack surface (S1192, S8513, S3358, S5799, S9073, S9100) Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * refactor: drop redundant HTTPError from except tuples (S5713); split a composite assert (S9073) HTTPError derives from URLError, so catching both is redundant; the behaviour is unchanged. Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV * refactor: single-return account handlers (S3516), split map_failure's HTTP branch (S3776), merge a literal (S5799) Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01CH5jAgaWUmCiXBJ3nKWdUV --------- Co-authored-by: Claude Opus 5 (1M context) <[email protected]>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
/initreplaced the scaffold seedCLAUDE.mdwith a runtime prompt grounded inthis repo, and the other three harness prompt files were brought in line with it —
each written for its own harness rather than restating the same text.
CLAUDE.mdQWEN.mdAGENTS.override.mdassociatelane — context only, identity stays in.pi/SYSTEM.mdAGENTS.colleague.md/ask-colleaguesecond mindREADME.mddrops the clone-and-rename-the-template instructions (this repo isthe clone) in favour of its own Status / CLI / Development sections.
Why
The scaffold left every doc describing
culture-agent-template, notsubstack-cli. Four concrete drifts are fixed:statistics and subscribers; no such noun groups are registered in
_commands/. Every file now states the scaffold status explicitly, andAGENTS.override.mdtells Pi's read-only lane to report that gap ratherthan infer an implementation from the project description.
uv run substack-cli whoami— but[project.scripts]installs the binary assubstack. Quickstart corrected;the binary-vs-prog-name mismatch is documented in all four harness files.
Left as a doc note deliberately — picking a canonical name is a code
change, not a docs one (see Follow-ups).
CLAUDE.md"the fullest write-up of the repo'sconventions" and pointed at a "Cloning this template" section it did not
contain — while
CLAUDE.mdwas still the/initseed placeholder.What's in CLAUDE.md now
Beyond commands: the cross-file CLI wiring that is not discoverable from any
single file — the
_CliArgumentParser.error()override that keeps argparsefailures inside the
error:/hint:contract, the_json_hintargv peek, theparser_classpropagation trap for nested nouns,CliErrorand the exit-codepolicy, the strict stdout/stderr split, and the fact that
tests/test_cli.pywalks
known_paths()so an uncatalogued verb fails the suite. Plus a five-placechecklist for adding a Substack noun, and the
_PROMPT_FILE(recognition) vs_RESIDENT_PROMPT(health) distinction thattests/test_doctor_resident_prompt.pypins shut.
Verification
uv run pytest -n auto— 117 passed, 1 skipped (the known report-onlycross-repo gap in
test_harness_registries.py)uv run teken cli doctor . --strict— PASSuv run python scripts/harness-smoke.py --stage config --require config— 6/6 PASSpython3 scripts/scan-secrets.py— clean;markdownlint-cli2— 0 errorsdevex pr lint— no violations;devex pr deltarun (alignment-file rule):no sibling follow-up needed, this PR touches no vendored skill.
Behavioural check. Both interactive harnesses were asked "Who are you, and
what can you do?" in this tree and answered from their new files — Qwen Code
named the repo by its real domain and framed the Substack verbs as work to add;
Pi answered with
associate's non-coding bounds from.pi/SYSTEM.mdplus thescaffold status from
AGENTS.override.md, volunteering that the Substacksurface "doesn't exist on disk yet".
ask-colleague explorewas also run — but it works from a throwaway worktree atHEAD, so it read the pre-change docs. Independently, from a cold read, it
flagged the same four drifts this PR fixes. A
reviewof the committed diff isrunning and findings will be folded in before merge.
Follow-ups (not in this PR)
substackvssubstack-cli: decide which is canonical and align[project.scripts]with the argparseprog, the explain catalog,learn,and the tests. Documented here, not papered over.
whoami.py's hand-rolledculture.yamlparser stops at the first agent blockwith no warning (raised by colleague). Harmless today — this repo declares one
agent — but worth a look before a multi-agent clone hits it.
substack-cli (Claude)