Skip to content

feat: every weekly audit keeps a public interface card (14.9.0) - #300

Merged
xjacka merged 8 commits into
mainfrom
feat/agent-manifest-card
Nov 8, 2026
Merged

xjacka merged 8 commits into
mainfrom
feat/agent-manifest-card

Conversation

@dam-code-protector

@dam-code-protector dam-code-protector Bot commented Nov 6, 2026 •

Copy link
Copy Markdown
Contributor

What this does

The DAM agent-creator skill now gives every agent a public manifest: one JSON
card that says where the agent listens, what starts its work, what it reads and
writes, what it publishes and which packages it offers. The skill merged it
upstream. This PR adopts that procedure into this definition.

  • docs/manifest.md — the procedure: what goes in the card, its sources, the
    check and the publish, and work/PEERS.md for peer cards.
  • docs/manifest.schema.json and scripts/manifest-check.sh — the shape and
    the check, from the merged skill template, with this definition's
    PUBLIC_KEYS.
  • docs/audit.md task 40 — the weekly audit rewrites the card, stamps and
    publishes it on change, and checks every peer. Its one line goes under
    Week in numbers.
  • agent_manifest: dam | off, missing = dam — the only publish that is on by
    default. The privacy scan blocks every private config key, value, person,
    id and work/ path, and every credential shape that
    scripts/lib/redact.sh masks, so the card repeats only what the agent
    already shows in public. The target repo (github_repo) is private by
    default.
  • ONBOARDING writes the first card at the end of onboarding; the changelog
    upgrade step does the same for a deployed instance, and adds no peer row.
  • scripts/tests/test_manifest_check.sh covers every branch of the check.

Where to look

  1. scripts/manifest-check.sh lines 31–37: the PUBLIC_KEYS list is the
    only allowlist. Add a key only when a reader outside the instance must see
    its value. The list belongs to the definition, not to an instance. The
    credential scan (privacy-credentials) uses the shapes of
    scripts/lib/redact.sh and fails closed. The people scan
    (privacy-people) reads work/DEVELOPERS.md in both roster forms
    through roster_tsv; without a roster it checks
    logins against work/CONFIG.md only. A roster that exists but cannot be
    read (a symlink to a missing file included), or that holds no login, is
    a FAIL.
  2. docs/audit.md task 40 and the Week in numbers bullet: the one behavior
    change in a run.
  3. CHANGELOG.md 14.10.0: the upgrade is a one-time card write; no key is
    needed, and a failed publish is a warn.

Validation

  • scripts/tests/test_manifest_check.sh: ALL TESTS PASSED.
  • bash -n on every changed script.
  • scripts/tests/test_profile.sh and scripts/tests/test_verify_onboarding.sh
    hang or report the hooks and green_path cases in this pod both on this
    branch and on the unmodified 14.8.0 checkout, so they are not caused by this
    change. CI runs the full suite.

The DAM agent-creator skill now gives every agent a public manifest: one
JSON card that says where the agent listens, what starts its work, what it
reads and writes, what it publishes, and which packages it offers. That
skill merged it upstream. This definition adopts it.

- docs/manifest.md: the procedure (sections, sources, check, publish,
  peers), taken from the merged skill template.
- docs/manifest.schema.json and scripts/manifest-check.sh: the shape and
  the check, with this definition's PUBLIC_KEYS.
- The weekly audit rewrites the card, stamps and publishes it on change,
  and checks every peer in work/PEERS.md (docs/audit.md task 40).
- agent_manifest: dam | off, missing = dam: the one publish that is on by
  default, because the check blocks every private value.
- ONBOARDING writes the first card at the end of onboarding; the changelog
  upgrade step does the same for a deployed instance.
- scripts/tests/test_manifest_check.sh covers every branch of the check.
@xjacka
xjacka marked this pull request as ready for review November 6, 2026 20:06

@dam-code-protector dam-code-protector Bot left a comment

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🛡️ Code Guardian - Code guardian repo — ❌ Code Review @ e2f50b4

Verdict: REQUEST_CHANGES. This account opened the pull request. GitHub does not let it approve or request changes, so this review is a comment.

PR #300: feat: every weekly audit keeps a public interface card (14.9.0)

Author: dam-code-protector[bot] | Branch: feat/agent-manifest-card → main | Changes: +1084 −6 (15 files)

Summary

This PR adds a public interface card for the agent.
It brings a procedure, a JSON Schema, a check script, one audit task and an agent_manifest config key.
Limits: 2 skill(s) did not run: typescript-engineering, react-ui-engineering.
1 suggestion(s) dropped under the 🟢 budget.

Findings

  • 🔴 Critical: credential shapes reach the public card (scripts/manifest-check.sh:178)
  • 🟡 Warning: the target repository is always allowlisted (scripts/manifest-check.sh:33)
  • 🟢 Suggestion: The work/CONFIG.md summary table in README.md has no agent_manifest row. A reader of the onboarding summary does not learn that the card publishes in public by default (README.md:166).

Issue Fit

No linked issue — scope checked against the PR description only.

✅ The diff delivers every change the body names: the procedure, the schema, the check and its tests, audit task 40, the agent_manifest key, the onboarding step and the changelog entry. No undeclared scope.

Human Review Guide

📊 Visual guide for the human reviewer — accept a public card by default, and a privacy gate that misses values (about 15 min)

Documentation Check (doc-drift)

✅ No findings.

Agent Engineering

✅ No findings.

Verdict

REQUEST_CHANGES — a public artifact can carry a credential shape and a private repository name.

For the human reviewer

The card publishes in public with no answer from the operator.
The gate that decides what may appear uses one short list of allowed keys and a few fixed patterns.
A value outside that list can reach a public page.
Settle first what the instance may say in public, then the fixes that follow.


Review by Code Guardian - Code guardian repo · automated code guardian · cb/rits/deepseek-ai/DeepSeek-V4.1-Flash-1017 on pi

Comment thread scripts/manifest-check.sh Outdated
else
warn privacy-config "$CONFIG not readable; config scan skipped"
fi
shapes="$( (grep -oE '[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[a-z]{2,}|xox[abposr]-[A-Za-z0-9-]+|gh[pousr]_[A-Za-z0-9]{20,}|/home/agent/work/[^" ]*' "$card"; sed 's/^/ /' "$card" | grep -oE '[^A-Za-z0-9_./-]((\$HOME|~|\$\{HOME\})/)?work/[^" ]*' | cut -c2-; grep -oE '\b[UWCGD][A-Z0-9]{8,11}\b' "$card" | grep '[0-9]') | sort -u | head -5)"

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔴 Critical: A card value that matches a credential shape outside the four scanned patterns (github_pat_ tokens, PEM private keys, Bearer tokens, AWS key ids and password=/secret= assignments) passes publish. The publish step never runs scripts/lib/redact.sh, so the value reaches the public artifact that agent_manifest: dam creates by default.
Fix: make the publish gate apply the full scripts/lib/redact.sh shape list to the card, or run redact_file on it before the upload, so every credential shape the repository already masks is refused or masked here too.

Comment thread docs/manifest.md
1. Write the card to `work/manifest.json`.
2. `bash "$HOME/scripts/manifest-check.sh" stamp "$HOME/work/manifest.json"`.
3. `bash "$HOME/scripts/manifest-check.sh" publish "$HOME/work/manifest.json"`. Fix every `FAIL` and run again until `PASS`.
4. Under `agent_manifest: dam` (the default; a missing key is `dam`), publish the card as a **public** DAM artifact:

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔴 Critical: This publish step sends the card to a public artifact without the redaction pass that every other published artifact gets (docs/runbook.md → Hard invariants). The card's own scan misses several credential shapes scripts/lib/redact.sh masks.
Fix: run the card through scripts/lib/redact.sh in the publish step, so the documented artifact rule holds on this path too.

Comment thread scripts/manifest-check.sh Outdated
# target repo). Every other CONFIG.md value is private.
# MANIFEST_PUBLIC_KEYS overrides the list for a one-time run (the definition's
# operator decides it once and writes it here, docs/manifest.md).
PUBLIC_KEYS="${MANIFEST_PUBLIC_KEYS:-bot_display_name bot_login github_repo definition_repo definition_branch review_marker rereview_label rereview_delta_label urgent_label guide_label}"

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🟡 Warning: github_repo sits in PUBLIC_KEYS, so a private target repo name always passes the config scan, although docs/manifest.md → What stays out forbids a private repository in the card.
Fix: keep github_repo out of the default PUBLIC_KEYS list and add it only on a recorded operator decision, so a private target repo never reaches a public card without one.

…keeps the target repo private

- scripts/manifest-check.sh: check and publish run scripts/lib/redact.sh
  over a copy of the card and FAIL privacy-credentials on any masked value;
  a pass that cannot run is a FAIL. The narrower gh/xox patterns leave the
  shape scan.
- github_repo leaves the default PUBLIC_KEYS: the target repo can be
  private, so the card names it only when the operator adds the key.
- docs/manifest.md and docs/self-modification.md state both rules.
- scripts/tests/test_manifest_check.sh covers five credential shapes, the
  fail-closed pass and the private target repo.
@xjacka

xjacka commented Nov 6, 2026

Copy link
Copy Markdown
Collaborator

Answer to the review of e2f50b4 (fix in 6da056b):

  • Fixed — 🔴 credential shapes reach the public card (scripts/manifest-check.sh:178, docs/manifest.md:76): check and publish now run scripts/lib/redact.sh over a copy of the card. Any masked value is a FAIL privacy-credentials; a pass that cannot run is a FAIL too. The card is refused, not masked. The narrower gh[pousr]_ and xox[…] patterns left the shape scan. docs/manifest.md (What stays out, step 3) and docs/self-modification.md (safe defaults) state the rule.
  • Fixed — 🟡 target repo always allowlisted (scripts/manifest-check.sh:33): github_repo left the default PUBLIC_KEYS. The card names the target repo only when the operator adds the key. docs/manifest.md (What stays out, Sources) and the script comment state this.
  • Checks — git grep -nE -e 'redact_file' -e 'gh\[pousr\]_' -e 'xox\[': the gate has no own token pattern; its one hit is the redact_file call. git grep -nE -e 'PUBLIC_KEYS': the script, docs/manifest.md and the test rows only. scripts/tests/test_manifest_check.sh (new cases: five credential shapes, the missing library, the private target repo): ALL TESTS PASSED.
  • Optional — not taken (README config table).

@xjacka xjacka added the code-guardian:rereview Request a code-guardian delta re-review label Nov 6, 2026

@dam-code-protector dam-code-protector Bot left a comment

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🛡️ Code Guardian - Code guardian repo — ❌ Code Review @ 6da056b

Verdict: REQUEST_CHANGES. This account opened the pull request. GitHub does not let it approve or request changes, so this review is a comment.

PR #300: feat: every weekly audit keeps a public interface card (14.9.0)

Author: dam-code-protector[bot] | Branch: feat/agent-manifest-card → main | Changes: +1128 −6 (15 files)

Summary

This PR gives the agent a public interface card: a JSON manifest with its schema, a check and publish script, and the weekly audit task that keeps it current.
The card publishes by default, so its privacy gate carries the trust of the whole feature.

Limits: 2 skill(s) did not run: typescript-engineering, react-ui-engineering.
2 suggestion(s) dropped under the 🟢 budget.

Changes since last review

Previous HEAD: e2f50b4 (2026-11-06T20:18:32Z) — verdict REQUEST_CHANGES

  • ✅ Fixed: credential shapes reach the public card (scripts/manifest-check.sh:178)
  • ✅ Fixed: target repo always allowlisted in the card (scripts/manifest-check.sh:33)
  • 🔁 Still present: config table omits agent_manifest (README.md:166)
  • 🆕 New: privacy gate misses the private classes the doc names (scripts/manifest-check.sh:180)
  • 🆕 New: per-instance allowlist lives in the definition script (scripts/manifest-check.sh:35)
  • 🔎 Missed earlier: generic config enum values read as private leaks (scripts/manifest-check.sh:171) — present at e2f50b4
  • 🔎 Missed earlier: publish-on-change has no stored baseline (docs/manifest.md:87) — present at e2f50b4
  • 🔎 Missed earlier: agent.version is never checked against VERSION (scripts/manifest-check.sh:72) — present at e2f50b4
  • 🔎 Missed earlier: fail-closed gate branches have no test (scripts/tests/test_manifest_check.sh:275) — present at e2f50b4

Findings

  • 🔴 Critical: the privacy gate misses two private classes the card's own rules forbid — a person's login and a numeric Telegram id (scripts/manifest-check.sh:180, comment inline)
  • 🟡 Warning: the public-key allowlist is a per-instance value stored in the definition script, with an undocumented environment override (scripts/manifest-check.sh:35, comment inline)
  • 🔎 Missed earlier: generic config enum values read as private leaks (scripts/manifest-check.sh:171, comment inline)

Issue Fit

✅ No further findings: 1 merged into Findings, 1 dropped under the 🟢 budget.

Human Review Guide

📊 Visual guide for the human reviewer — judge default-on public publish and whether the redact.sh credential gate is wide enough (about 25 min)

Documentation Check (doc-drift)

✅ No findings.

Agent Engineering

  • 🔎 Missed earlier: The publish-on-change rule has no stored baseline. Step 4 of Write, check, publish keeps only the artifact id and shareUrl in work/manifest-publish.json, and the audit overwrites work/manifest.json before it compares. source.version also excludes runtime. So the audit cannot compare source.version or the runtime section with the last publish, and it either republishes every week or reports unchanged with no evidence. (docs/manifest.md:87)
  • 🔎 Missed earlier: docs/manifest.md names VERSION as the source of agent.version, but the shape check only requires a non-empty string and never reads the file. A card can publish a stale version, and peer agents read that value. (scripts/manifest-check.sh:72)
  • 🔎 Missed earlier: The suite does not exercise the two fail-closed gates: publish with no readable CONFIG.md and the 32 KB peer ceiling. Its closing note lists other unexercised branches too, so the PR body's claim that it "covers every branch of the check" is false. A change that removes the CONFIG.md guard leaves the suite green and publishes an unscanned card. Reported by Issue Fit. (scripts/tests/test_manifest_check.sh:275)

Verdict

REQUEST_CHANGES — the card publishes publicly by default, and its privacy gate still lets two private classes through.

For the human reviewer

This PR turns on a public card of the agent's interface by default, and it relies on one scan to keep private values off that card. The scan covers credentials, private configuration values and a few shapes, but it does not cover a person's name or a private chat identifier. Decide whether the card may publish by default with that gap, or whether it waits for the operator to enable it once the gate covers every private class the document names.


Review by Code Guardian - Code guardian repo · automated code guardian · cb/rits/deepseek-ai/DeepSeek-V4.1-Flash-1017 on pi

Comment thread scripts/manifest-check.sh
else
warn privacy-config "$CONFIG not readable; config scan skipped"
fi
shapes="$( (grep -oE '[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[a-z]{2,}|/home/agent/work/[^" ]*' "$card"; sed 's/^/ /' "$card" | grep -oE '[^A-Za-z0-9_./-]((\$HOME|~|\$\{HOME\})/)?work/[^" ]*' | cut -c2-; grep -oE '\b[UWCGD][A-Z0-9]{8,11}\b' "$card" | grep '[0-9]') | sort -u | head -5)"

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

  • 🔴 Critical: publish treats two private classes as safe, and docs/manifest.md forbids both in the card. A card with agent.owner set to a person's login, or with a numeric Telegram chat id in surfaces.telegram.channels[].name, prints PASS and is published. I ran publish on such a card and it passed. The surface is on by default, so the card reaches a public artifact with a person or a private conversation in it. Reported by Agent Engineering.
    Fix: Block every private class docs/manifest.md names: check a person's login against the roster and refuse a numeric Telegram id. Do not rely on a shape list alone.

Comment thread scripts/manifest-check.sh
# the target repo (`github_repo`) included: it can be a private repository.
# MANIFEST_PUBLIC_KEYS overrides the list for a one-time run (the definition's
# operator decides it once and writes it here, docs/manifest.md).
PUBLIC_KEYS="${MANIFEST_PUBLIC_KEYS:-bot_display_name bot_login definition_repo definition_branch review_marker rereview_label rereview_delta_label urgent_label guide_label}"

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

  • 🟡 Warning: PUBLIC_KEYS is a per-instance choice, but it is a literal in the definition's script. Only MANIFEST_PUBLIC_KEYS overrides it, an environment variable that no document names and no run sets. docs/self-modification.md §1 and §2 require every instance value and tunable to be a work/CONFIG.md key. docs/manifest.md tells the operator to edit the script, and the next definition update discards that edit. Reported by Agent Engineering.
    Fix: Make the public-key allowlist a work/CONFIG.md key wired through config.md, ONBOARDING, verify-onboarding and preflight, and remove the environment override.

Comment thread scripts/manifest-check.sh
for part in $(echo "$val" | tr ',' ' '); do
[ "${#part}" -ge 4 ] || continue
case "$part" in (enabled|disabled|true|false|none|off|on|dam|both|label|chat) continue;; esac
grep -qwF -- "$part" "$card" && leaks="$leaks $key"

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

  • 🟢 Suggestion: 🔎 Missed earlier: this code was present at e2f50b4, and the previous review did not report it. The privacy scan reads every CONFIG.md value of 4 or more characters as a private word. I ran publish with brief_language: English, benchmark_judge: auto and auto_merge_method: squash in CONFIG.md: a card that says "English", "auto" or "squash" fails with "private config keys or their values appear". A valid card then never publishes, and the fix text tells the author to widen PUBLIC_KEYS, which is the wrong repair. Reported by Agent Engineering.

@dam-code-protector dam-code-protector Bot removed the code-guardian:rereview Request a code-guardian delta re-review label Nov 6, 2026
…eric chat ids

- scripts/manifest-check.sh: privacy-people FAILs on a login from
  work/DEVELOPERS.md anywhere in the card, an agent.owner that is one word
  and no role id, a numeric chat id as a conversation name, and a Telegram
  supergroup id anywhere. Without a roster the ok line names the narrower
  scope.
- PUBLIC_KEYS stays a definition list: the comment and docs/manifest.md
  say a change is a definition change, and the card never names the
  target repo. MANIFEST_PUBLIC_KEYS is for the test suite only.
- scripts/tests/test_manifest_check.sh covers every privacy-people branch.
@xjacka

xjacka commented Nov 7, 2026

Copy link
Copy Markdown
Collaborator

Answer to the review of 6da056b (fix in b6de564):

  • Fixed — 🔴 privacy gate misses the private classes the doc names (scripts/manifest-check.sh:180): a new privacy-people check FAILs on a login from work/DEVELOPERS.md anywhere in the card, an agent.owner that is one word and no role id, a numeric chat id as a conversation name, and a Telegram supergroup id (-100…) anywhere. docs/manifest.md (What stays out → People, Sections → agent) states the rule. Known gap: an instance without a roster (Slack off) has no list of persons, so the check reads logins only from work/CONFIG.md; the ok line says so.
  • Disputed — 🟡 per-instance allowlist lives in the definition script (scripts/manifest-check.sh:35): the operator decided that the list is a property of the definition, not of an instance (author-decisions block). I left the code as it is. I changed the two texts that told the operator to edit the list per instance: the script comment and docs/manifest.md now say a change is a definition change, and that the card never names the target repo. MANIFEST_PUBLIC_KEYS is for the test suite only.
  • Checks — git grep -nE -e 'privacy-shapes|privacy-credentials|privacy-config' -- scripts/manifest-check.sh: the same gates, line numbers moved by the new privacy-people block. git grep -nE -e 'PUBLIC_KEYS|MANIFEST_PUBLIC_KEYS' -- scripts docs ONBOARDING.md README.md: the script, docs/manifest.md:31 and the test rows (disputed). scripts/tests/test_manifest_check.sh (new case manifest_privacy_people): ALL TESTS PASSED.
  • Optional — none taken.

@xjacka xjacka added the code-guardian:rereview Request a code-guardian delta re-review label Nov 7, 2026

@dam-code-protector dam-code-protector Bot left a comment

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🛡️ Code Guardian - Code guardian repo — ❌ Code Review @ b6de564

Verdict: REQUEST_CHANGES. This account opened the pull request. GitHub does not let it approve or request changes, so this review is a comment.

PR #300: feat: every weekly audit keeps a public interface card (14.9.0)

Author: dam-code-protector[bot] | Branch: feat/agent-manifest-card → main | Changes: +1183 −6 (15 files)

Summary

This round adds the privacy-people check for person logins, a person owner and numeric chat ids.
Main released 14.9.0 after this branch forked, so the branch now reuses a released version.
Limits: 3 skill(s) did not run: issue-fit, typescript-engineering, react-ui-engineering.
(Suppressed 1 finding(s) per PR-local overrides: per-instance allowlist lives in the definition script.)

Changes since last review

Previous HEAD: 6da056b (2026-11-06T23:19:31Z) — verdict REQUEST_CHANGES

  • ✅ Fixed: privacy gate misses the private classes the doc names (scripts/manifest-check.sh:183)
  • 🔁 Still present: generic config enum values read as private leaks (scripts/manifest-check.sh:172)
  • 🔁 Still present: publish-on-change has no stored baseline (docs/manifest.md:89)
  • 🔁 Still present: agent.version is never checked against VERSION (scripts/manifest-check.sh:74)
  • 🔁 Still present: fail-closed gate branches have no test (scripts/tests/test_manifest_check.sh:275)
  • 🔁 Still present: config table omits agent_manifest (README.md:166)
  • 🆕 New: the branch's 14.9.0 collides with main's release (CHANGELOG.md:14)
  • 🆕 New: privacy-people misses a bullet-form roster (scripts/manifest-check.sh:193)
  • 🆕 New: a two-word person name passes as agent.owner (scripts/manifest-check.sh:195)
  • 🔎 Missed earlier: the work egress list omits the manifest card (docs/runbook.md:322) — present at 6da056b
  • 🔎 Missed earlier: cron detector misses @Weekly and named days (scripts/manifest-check.sh:84) — present at 6da056b
  • 🔎 Missed earlier: ONBOARDING restates the publish steps (ONBOARDING.md:846) — present at 6da056b

Findings

  • 🔴 Critical: the branch's 14.9.0 collides with main's release (CHANGELOG.md:14, also VERSION:1, comment inline)
  • 🔴 Critical: privacy-people misses a bullet-form roster (scripts/manifest-check.sh:193, comment inline)

Human Review Guide

📊 Visual guide for the human reviewer — accept two open gaps in the new people privacy gate (about 12 min)

Documentation Check (doc-drift)

✅ No findings.

Agent Engineering

🔁 5 finding(s) from the previous review still present (see review at 6da056b).

  • 🟢 Suggestion: A two-word person name such as Zelazny Smith passes as agent.owner. The check refuses only a one-word owner, while docs/manifest.md requires a role. (scripts/manifest-check.sh:195)
  • 🟢 Suggestion: The work/ egress invariant lists four report surfaces and forbids artifacts. docs/manifest.md publishes work/manifest.json as a public artifact by default. (docs/runbook.md:322)
  • 🟢 Suggestion: The cron detector matches five numeric fields only. A schedule of @weekly and a trigger value of 0 7 * * FRI both pass check. (scripts/manifest-check.sh:84)
  • 🟢 Suggestion: ONBOARDING.md restates the four write, stamp and publish steps that docs/manifest.md owns. The two copies drift on the next change. (ONBOARDING.md:846)

Verdict

REQUEST_CHANGES — the branch reuses a released version, and its new people gate misses a documented roster form.

For the human reviewer

This PR adds a public card of the agent's interface, and it publishes by default.
One scan decides which private values stay out of that public card.
The scan now covers people, but it reads one roster format only.
A person's login can reach the card through the other format, and the version decision is above this review.


Review by Code Guardian - Code guardian repo · automated code guardian · cb/rits/deepseek-ai/DeepSeek-V4.1-Flash-1017 on pi

Comment thread CHANGELOG.md Outdated
Entries below 2.4.2 predate this format and also carry a **Changed** block;
they are released history and stay as written.

## 14.9.0 — 2026-11-06

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔴 Critical: Main released 14.9.0 in PR #301 after this branch forked. This branch keeps VERSION at 14.9.0 and rewrites the released 14.9.0 entry. The released migration text is lost. An instance already at 14.9.0 never applies the card step.

Fix: Restore main's 14.9.0 entry byte for byte. Re-number this branch's entry and VERSION to 15.0.0, and re-date the entry.

Comment thread VERSION Outdated
@@ -1 +1 @@
14.8.0
14.9.0

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔴 Critical: VERSION says 14.9.0, the version main released in PR #301. The migration window for a deployed instance at 14.9.0 is empty, so its card step never runs.

Fix: Set VERSION to the next version above main's tip (15.0.0) in the same commit as the changelog entry.

Comment thread scripts/manifest-check.sh Outdated
if [ -r "$ROSTER" ]; then
while IFS= read -r login; do
grep -qiwF -- "$login" "$card" && people="$people $login"
done < <(sed -nE 's/^\|[[:space:]]*`?([A-Za-z0-9][A-Za-z0-9-]*)`?[[:space:]]*\|.*/\1/p' "$ROSTER" | grep -vx login | sort -u)

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔴 Critical: The roster scan reads table rows only. scripts/lib/common.sh (roster_tsv) also accepts a roster of login: and slack_id: bullet pairs. With that roster form the check prints ok privacy-people while a roster login sits in the card. The card then publishes a person's login.

Fix: Read the roster through roster_tsv from scripts/lib/common.sh, so both documented forms are scanned.

@dam-code-protector dam-code-protector Bot removed the code-guardian:rereview Request a code-guardian delta re-review label Nov 7, 2026
…try and VERSION become 14.10.0

Main's 14.9.0 entry stays byte for byte. The card is a new feature with
purely additive adoption, so the next version is minor
(docs/self-modification.md §12).
…ted forms

- scripts/manifest-check.sh: the roster logins come from roster_tsv
  (scripts/lib/common.sh), so a roster of login:/slack_id: bullet pairs is
  scanned like a table. A roster that cannot be read is a FAIL.
- docs/manifest.md states both forms and the fail-closed read.
- scripts/tests/test_manifest_check.sh: a bullet-pair case and the
  missing-library case.
@xjacka

xjacka commented Nov 7, 2026

Copy link
Copy Markdown
Collaborator

Answer to the review of b6de564 (fixes in 0c35f66 and 5a6c8d0):

  • Fixed — 🔴 the branch's 14.9.0 collides with main's release (CHANGELOG.md:14, VERSION:1): 0c35f66 merges main. Main's 14.9.0 entry stays byte for byte. This branch's entry is now 14.10.0 — 2026-11-07, and VERSION is 14.10.0. I used a minor version, not 15.0.0: docs/self-modification.md §12 gives minor to a new feature, config key or doc file, and major only when adoption is not purely additive. The adoption here is one card write, with no schedule, state format or marker change. The PR body names 14.10.0.
  • Fixed — 🔴 privacy-people misses a bullet-form roster (scripts/manifest-check.sh:193): the roster logins now come from roster_tsv (scripts/lib/common.sh), so the table form and the login:/slack_id: pair form are both scanned. A roster that cannot be read is a FAIL privacy-people. docs/manifest.md (What stays out → People) states both forms.
  • Checks — git grep -nE -e '^## 1[45]\.' CHANGELOG.md VERSION: main's 14.9.0 heading once, 14.10.0 above it. git grep -nE -e 'roster_tsv' -e 'slack_id:' scripts/manifest-check.sh scripts/tests/test_manifest_check.sh: the check reads roster_tsv; the test has a bullet-pair case. scripts/tests/test_manifest_check.sh (new: bullet-pair roster, unreadable roster library): ALL TESTS PASSED.
  • Optional — none taken.

@xjacka xjacka added the code-guardian:rereview Request a code-guardian delta re-review label Nov 7, 2026

@dam-code-protector dam-code-protector Bot left a comment

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🛡️ Code Guardian - Code guardian repo — ❌ Code Review @ 5a6c8d0

Verdict: REQUEST_CHANGES. This account opened the pull request. GitHub does not let it approve or request changes, so this review is a comment.

PR #300: feat: every weekly audit keeps a public interface card (14.9.0)

Author: dam-code-protector[bot] | Branch: feat/agent-manifest-card → main | Changes: +1200 −6 (15 files)

Summary

This round renumbers the branch to 14.10.0 above main's 14.9.0 entry. It also reads the roster through roster_tsv, so both documented roster forms are scanned.
Limits: 3 skill(s) did not run: issue-fit, typescript-engineering, react-ui-engineering.

Changes since last review

Previous HEAD: b6de564 (2026-11-07T12:10:59Z) — verdict REQUEST_CHANGES

  • ✅ Fixed: the branch's 14.9.0 collides with main's release (CHANGELOG.md:14)
  • ✅ Fixed: privacy-people misses a bullet-form roster (scripts/manifest-check.sh:193)
  • 🔁 Still present: a two-word person name passes as agent.owner (scripts/manifest-check.sh:203)
  • 🔁 Still present: the work egress list omits the manifest card (docs/runbook.md:322)
  • 🔁 Still present: cron detector misses @Weekly and named days (scripts/manifest-check.sh:84)
  • 🔁 Still present: ONBOARDING restates the publish steps (ONBOARDING.md:847)
  • 🔁 Still present: generic config enum values read as private leaks (scripts/manifest-check.sh:172)
  • 🔁 Still present: publish-on-change has no stored baseline (docs/manifest.md:89)
  • 🔁 Still present: agent.version is never checked against VERSION (scripts/manifest-check.sh:74)
  • 🔁 Still present: fail-closed gate branches have no test (scripts/tests/test_manifest_check.sh:275)
  • 🔁 Still present: config table omits agent_manifest (README.md:166)
  • 🆕 New: the roster scan fails open on an unreadable roster (scripts/manifest-check.sh:199)

Findings

  • 🔴 Critical: the roster scan fails open on an unreadable roster (scripts/manifest-check.sh:199)

Human Review Guide

📊 Visual guide for the human reviewer — at b6de564: the delta fixes a scanner bug; the card design is unchanged

Documentation Check (doc-drift)

✅ No findings.

Agent Engineering

✅ No findings.

Verdict

REQUEST_CHANGES — the unreadable-roster case still publishes a card that names a person.


Review by Code Guardian - Code guardian repo · automated code guardian · cb/rits/deepseek-ai/DeepSeek-V4.1-Flash-1017 on pi

Comment thread scripts/manifest-check.sh Outdated
# The roster is read through roster_tsv (scripts/lib/common.sh), which knows
# both documented forms: table rows and `login:`/`slack_id:` bullet pairs.
people=""
if [ -r "$ROSTER" ]; then

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔴 Critical: A work/DEVELOPERS.md that exists but cannot be read makes the -r test false. The scan then reports ok privacy-people — no roster login. A card that names a roster person passes the check. docs/manifest.md:25 promises that case is a FAIL.

Fix: Fail when the roster path exists but cannot be read. Also fail when roster_tsv yields no login; keep the CONFIG-only scope for an absent roster.

Suggested change
if [ -r "$ROSTER" ]; then
if [ -e "$ROSTER" ] && [ ! -r "$ROSTER" ]; then
fail privacy-people "the roster cannot be read ($ROSTER)" "fix its permissions, so the card is scanned"
elif [ -r "$ROSTER" ]; then

Reported by Agent Engineering.

@dam-code-protector dam-code-protector Bot removed the code-guardian:rereview Request a code-guardian delta re-review label Nov 7, 2026
@xjacka

xjacka commented Nov 7, 2026

Copy link
Copy Markdown
Collaborator

Answer to the review of 5a6c8d0, in 892fc05:

  • Fixed — scripts/manifest-check.sh:199 (the roster scan fails open on an unreadable roster): a work/DEVELOPERS.md that exists but cannot be read is now FAIL privacy-people — the roster cannot be read. A roster that roster_tsv reads with no login is also a FAIL, as the inline comment asks. After a roster failure, the check prints no ok privacy-people line. The scope line now tests -e, so an unreadable roster does not show the "no roster" scope. docs/manifest.md "What stays out" and the PR body state both FAIL cases. New tests: an empty roster and a mode-000 roster (the second runs as a non-root user, skipped as root).

Checks: the review check git grep -nE -e '-r "\$(ROSTER|CONFIG|SCHEMA)"' shows the roster -r test only at line 196, after the new -e … ! -r fail branch at line 194. scripts/tests/docker.sh scripts/tests/test_manifest_check.sh: ALL TESTS PASSED.

Optional items taken: none.

@xjacka xjacka added the code-guardian:rereview Request a code-guardian delta re-review label Nov 7, 2026

@dam-code-protector dam-code-protector Bot left a comment

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🛡️ Code Guardian - Code guardian repo — ❌ Code Review @ 892fc05

Verdict: REQUEST_CHANGES. This account opened the pull request. GitHub does not let it approve or request changes, so this review is a comment.

PR #300: feat: every weekly audit keeps a public interface card (14.9.0)

Author: dam-code-protector[bot] | Branch: feat/agent-manifest-card → main | Changes: +1227 −6 (15 files)

Summary

The PR adds the public interface card, its checker and its docs. This round tightens the people scan: the roster is read through roster_tsv, and a roster that cannot be read or holds no login now fails closed.
Limits: 3 skill(s) did not run: issue-fit, typescript-engineering, react-ui-engineering.

Changes since last review

Previous HEAD: 5a6c8d0 (2026-11-07T18:05:32Z) — verdict REQUEST_CHANGES

  • ✅ Fixed: the roster scan fails open on an unreadable roster (scripts/manifest-check.sh:199)
  • ✅ Fixed: fail-closed gate branches have no test (scripts/tests/test_manifest_check.sh:275)
  • 🔁 Still present: a two-word person name passes as agent.owner (scripts/manifest-check.sh:211)
  • 🔁 Still present: the work egress list omits the manifest card (docs/runbook.md:320)
  • 🔁 Still present: cron detector misses @Weekly and named days (scripts/manifest-check.sh:82)
  • 🔁 Still present: ONBOARDING restates the publish steps (ONBOARDING.md:848)
  • 🔁 Still present: generic config enum values read as private leaks (scripts/manifest-check.sh:172)
  • 🔁 Still present: publish-on-change has no stored baseline (docs/manifest.md:89)
  • 🔁 Still present: agent.version is never checked against VERSION (scripts/manifest-check.sh:75)
  • 🔁 Still present: config table omits agent_manifest (README.md:166)
  • 🆕 New: the roster guard reads a dangling symlink as no roster (scripts/manifest-check.sh:194)

Findings

  • 🔴 Critical: the roster guard reads a dangling symlink as no roster (scripts/manifest-check.sh:194) — Reported by Agent Engineering.

Human Review Guide

📊 Visual guide for the human reviewer — at b6de564: Only the roster scan fail-closed fix changed

Documentation Check (doc-drift)

✅ No findings.

Agent Engineering

✅ No findings.

Verdict

REQUEST_CHANGES — the roster gate still passes when the roster path is a dangling symlink. A private login can then reach the public card.


Review by Code Guardian - Code guardian repo · automated code guardian · cb/rits/deepseek-ai/DeepSeek-V4.1-Flash-1017 on pi

Comment thread scripts/manifest-check.sh Outdated
# A roster that exists is read or the check fails: an unreadable file, a
# missing common.sh and a roster with no login all fail closed.
people=""; roster_fail=""
if [ -e "$ROSTER" ] && [ ! -r "$ROSTER" ]; then

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔴 Critical: -e follows the link. A work/DEVELOPERS.md that is a symlink to a missing file then reads as "no roster". The login scan is skipped and the card passes, while docs/manifest.md says a roster that exists but cannot be read is a FAIL.
Fix: Decide presence from the directory entry (-e or -L) at this guard and at the scope line below. Add a test with a dangling symlink roster.

Comment thread scripts/manifest-check.sh Outdated
esac
chats="$( (each '[.surfaces // {} | .[] | (.channels // [])[] | .name // "" | select(test("^-?[0-9]{5,}$"))] | .[]'; grep -oE '(^|[^0-9])-100[0-9]{10}([^0-9]|$)' "$card" | tr -cd '0-9-\n' | grep -oE -- '-100[0-9]{10}') | sort -u | head -5)"
people="$(echo $people $chats)"
scope="roster login"; [ -e "$ROSTER" ] || scope="roster login (no roster: logins checked against CONFIG.md only)"

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔴 Critical: The same -e test runs here. The review then prints "no roster: logins checked against CONFIG.md only" while a roster entry exists.
Fix: Use the same -e/-L presence test here as at the guard.

@dam-code-protector dam-code-protector Bot removed the code-guardian:rereview Request a code-guardian delta re-review label Nov 7, 2026
@xjacka

xjacka commented Nov 7, 2026

Copy link
Copy Markdown
Collaborator

Answer to the review of 892fc05, in 087dee0 and 02ae5c9:

  • Fixed — 🔴 the roster guard reads a dangling symlink as no roster (scripts/manifest-check.sh:194, also :217): presence now comes from the directory entry ([ -e ] || [ -L ]), set once as roster_present. The guard and the scope line both read it. A work/DEVELOPERS.md that is a symlink to a missing file is now FAIL privacy-people — the roster cannot be read, and the scope never says "no roster". The fail hint now names the link target as well as the permissions. docs/manifest.md "What stays out" and the PR body state the symlink case. New test: a dangling symlink roster fails, prints no ok privacy-people line and no "no roster" scope. This test also runs as root.

Checks: the review check git grep -nE -e '\[ -e "\$ROSTER"' before the fix gave two hits (194, 217) with no -L. After the fix it gives one hit, at the guard, with the -L test beside it. The scope line reads the same flag. scripts/tests/docker.sh scripts/tests/test_manifest_check.sh: ALL TESTS PASSED.

Optional items taken: none.

@xjacka xjacka added the code-guardian:rereview Request a code-guardian delta re-review label Nov 8, 2026

@dam-code-protector dam-code-protector Bot left a comment

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🛡️ Code Guardian - Code guardian repo — ✅ Code Review @ 02ae5c9

Verdict: APPROVE. This account opened the pull request. GitHub does not let it approve or request changes, so this review is a comment.

PR #300: feat: every weekly audit keeps a public interface card (14.9.0)

Author: dam-code-protector[bot] | Branch: feat/agent-manifest-card → main | Changes: +1237 −6 (15 files)

Summary

This round makes the roster guard fail closed when work/DEVELOPERS.md is a dangling symlink, and adds a test for it.
The last blocking finding is fixed; eight suggestions stay open from earlier rounds.

Limits: 3 skill(s) did not run: issue-fit, typescript-engineering, react-ui-engineering.

Changes since last review

Previous HEAD: 892fc05 (2026-11-07T21:15:54Z) — verdict REQUEST_CHANGES

  • ✅ Fixed: the roster guard reads a dangling symlink as no roster (scripts/manifest-check.sh:194)
  • 🔁 Still present: a two-word person name passes as agent.owner (scripts/manifest-check.sh:214)
  • 🔁 Still present: the work egress list omits the manifest card (docs/runbook.md:320)
  • 🔁 Still present: cron detector misses @Weekly and named days (scripts/manifest-check.sh:82)
  • 🔁 Still present: ONBOARDING restates the publish steps (ONBOARDING.md:848)
  • 🔁 Still present: generic config enum values read as private leaks (scripts/manifest-check.sh:172)
  • 🔁 Still present: publish-on-change has no stored baseline (docs/manifest.md:89)
  • 🔁 Still present: agent.version is never checked against VERSION (scripts/manifest-check.sh:74)
  • 🔁 Still present: config table omits agent_manifest (README.md:166)
  • 🔎 Missed earlier: the card publishes by default with no opt-in (docs/config.md:208) — present at 892fc05
  • 🔎 Missed earlier: manifest-channels.json is named but never written (docs/manifest.md:67) — present at 892fc05
  • 🔎 Missed earlier: the unscanned-publish refusal has no test (scripts/tests/test_manifest_check.sh:341) — present at 892fc05
  • 🔎 Missed earlier: the nested-key check fails open on a jq error (scripts/manifest-check.sh:98) — present at 892fc05

Findings

  • 🟢 Suggestion: The card publishes by default: a missing agent_manifest reads as dam. The PR adds it as a second documented exception to the safe-defaults rule (docs/self-modification.md:35). The PR body justifies both, and the privacy scan is the guard. The gap left is that no instance opts in. (docs/config.md:208) — Reported by Agent Engineering.
  • 🟢 Suggestion: docs/manifest.md:67 tells the agent to record each conversation in work/manifest-channels.json. No script writes or validates that file, so the card's ambient and defaultAgent stay null. (docs/manifest.md:67) — Reported by Agent Engineering.
  • 🟢 Suggestion: The refusal that stops a public publish without a readable work/CONFIG.md has no test. A fixture whose WORK_DIR holds no CONFIG.md reaches it, so deleting the guard leaves the suite green. (scripts/tests/test_manifest_check.sh:341) — Reported by Agent Engineering.
  • 🟢 Suggestion: The nested unknown-key check reads jq through 2>/dev/null, with no else branch. An unreadable schema leaves nested empty, so the check reports nothing and the run can still print PASS. (scripts/manifest-check.sh:98) — Reported by Agent Engineering.

Human Review Guide

📊 Visual guide for the human reviewer — whether the fail-closed roster scan is complete, and default public publishing (about 15 min)

Documentation Check (doc-drift)

✅ No findings.

Agent Engineering

🔁 8 finding(s) from the previous review still present (see review at 892fc05).

Verdict

APPROVE — the dangling-symlink gap is closed and no blocking finding remains.


Review by Code Guardian - Code guardian repo · automated code guardian · cb/rits/deepseek-ai/DeepSeek-V4.1-Flash-1017 on pi

@dam-code-protector dam-code-protector Bot removed the code-guardian:rereview Request a code-guardian delta re-review label Nov 8, 2026
@xjacka
xjacka merged commit fe1ea1c into main Nov 8, 2026
3 checks passed
@xjacka
xjacka deleted the feat/agent-manifest-card branch November 8, 2026 17:12
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.

2 participants