fix(review-pipeline): improved-from-#6 — narrow code-only successor #7

Merged

pdurlej merged 2 commits from codex/review-pipeline/improved-from-pr6 into main

2026-05-01 17:39:52 +02:00

pdurlej commented

2026-05-01 17:25:25 +02:00

Owner

Successor to closed/rejected #6

Per operator architectural decision (2026-05-01) + Oracle consult (session platform-meta-decisions-2026-05-01, gpt-5.5 high reasoning, fallback via codex CLI because Oracle MCP browser auth not set up in claude-orchestrator's environment — see operator INBOX for diagnostics):

"PR rejection as learning log" pattern is sound for AI-driven workflow when used for architectural / scope-creep / hygiene contamination — exactly this case.

Closes #6.

Removed (vs PR #6)

state/DISPATCH_BRIDGE.json — machine-local session state (4 reviewers convergent in canary #3)
state/canary/CANARY_PLAN.md Path C OpenCode credential-extraction (4 reviewers convergent: don't normalize credential scavenging)
state/reviews/PR-{5,6}/* review artifact JSONs (per reviewer feedback: cause self-referential confusion in subsequent reviews)
state/{AUDIT_LOG,OPERATOR_INBOX,STATUS_NOW} updates — these evolve naturally; not bundled into code PRs

Retained (the actual fixes from canary #1 + canary #2 + canary #3)

File	Change	Source
`_base.py`	parser iterates REVERSE, picks LAST verdict block	canary #1 BLOCKER, 3/3 tech convergent
`claude_cli.py`	stdin (NOT argv); `--bare` opt-in via `PLATFORMCTL_CLAUDE_BARE=1`	canary #1 (argv leak) + canary #2 (--bare killed OAuth)
`codex_cli.py`	`--ephemeral` flag added	tech-gpt opportunity
`zai.py`	default URL `api/coding/paas/v4` (Coding Plan); `/tmp` leak removed	canary #1 (high) + ops fix
`test_run_review.py`	regression test for parser bug class	canary #1 (high)

Changed (charter)

PLATFORM_CHARTER.md §3 — added "Learning patterns" subsection with two charter rules (PR rejection as learning log + Oracle escalation discipline). Per Oracle consult.

Tests

60/60 green:

37 in test_run_review.py (incl. new regression test that would FAIL pre-fix)
14 consolidator
4 negative-control
5 platformctl smoke

Linkage

Closes #6
Builds on canary #1, #2, #3 findings (in PR #6 history + operator INBOX narrative)
First exercise of "PRs as learning log" pattern enshrined in this PR's charter update

🤖 Generated with Claude Code

## Successor to closed/rejected #6 Per operator architectural decision (2026-05-01) + Oracle consult (session `platform-meta-decisions-2026-05-01`, gpt-5.5 high reasoning, fallback via codex CLI because Oracle MCP browser auth not set up in claude-orchestrator's environment — see operator INBOX for diagnostics): > "PR rejection as learning log" pattern is sound for AI-driven workflow when used for architectural / scope-creep / hygiene contamination — exactly this case. Closes #6. ## Removed (vs PR #6) - `state/DISPATCH_BRIDGE.json` — machine-local session state (4 reviewers convergent in canary #3) - `state/canary/CANARY_PLAN.md` Path C OpenCode credential-extraction (4 reviewers convergent: don't normalize credential scavenging) - `state/reviews/PR-{5,6}/*` review artifact JSONs (per reviewer feedback: cause self-referential confusion in subsequent reviews) - `state/{AUDIT_LOG,OPERATOR_INBOX,STATUS_NOW}` updates — these evolve naturally; not bundled into code PRs ## Retained (the actual fixes from canary #1 + canary #2 + canary #3) | File | Change | Source | |---|---|---| | `_base.py` | parser iterates REVERSE, picks LAST verdict block | canary #1 BLOCKER, 3/3 tech convergent | | `claude_cli.py` | stdin (NOT argv); `--bare` opt-in via `PLATFORMCTL_CLAUDE_BARE=1` | canary #1 (argv leak) + canary #2 (--bare killed OAuth) | | `codex_cli.py` | `--ephemeral` flag added | tech-gpt opportunity | | `zai.py` | default URL `api/coding/paas/v4` (Coding Plan); `/tmp` leak removed | canary #1 (high) + ops fix | | `test_run_review.py` | regression test for parser bug class | canary #1 (high) | ## Changed (charter) - `PLATFORM_CHARTER.md` §3 — added "Learning patterns" subsection with two charter rules (PR rejection as learning log + Oracle escalation discipline). Per Oracle consult. ## Tests 60/60 green: - 37 in `test_run_review.py` (incl. new regression test that would FAIL pre-fix) - 14 consolidator - 4 negative-control - 5 platformctl smoke ## Linkage - Closes #6 - Builds on canary #1, #2, #3 findings (in PR #6 history + operator INBOX narrative) - First exercise of "PRs as learning log" pattern enshrined in this PR's charter update 🤖 Generated with [Claude Code](https://claude.com/claude-code)

pdurlej added 1 commit

2026-05-01 17:25:25 +02:00

fix(review-pipeline): improved-from-#6 — narrow code-only successor 2c1db04714

Successor PR to closed/rejected #6 (convergent reviewer feedback: split
before merge — 3-4 reviewers convergent on scope creep + machine-local
state + credential extraction in canonical doc).

Per Oracle consult (2026-05-01, gpt-5.5 high reasoning, session
`platform-meta-decisions-2026-05-01`): "PR rejection as learning log"
pattern is sound for AI-driven workflow when used for architectural /
scope-creep / hygiene contamination — exactly this case.

## Removed (vs PR #6)

- `state/DISPATCH_BRIDGE.json` — machine-local session state (4 reviewers
  convergent: shouldn't be committed; goes to .runtime/ or .gitignore)
- `state/canary/CANARY_PLAN.md` Path C OpenCode credential-extraction
  documentation (4 reviewers convergent: don't normalize credential
  scavenging; goes to private runbook)
- `state/reviews/PR-{5,6}/*` review artifact JSONs (per reviewer
  feedback: cause self-referential confusion in subsequent reviews;
  goes to separate archive repo or out-of-repo storage)
- `state/{AUDIT_LOG,OPERATOR_INBOX,STATUS_NOW}.md` updates — these
  evolve naturally on main; will accumulate organically

## Retained (the actual fixes from canary #1 + #2 + #3 findings)

- `_base.py` parser: REVERSE iteration picks LAST verdict-containing
  block (canary #1 BLOCKER fix, 3/3 tech reviewers convergent)
- `claude_cli.py`: stdin (NOT argv) for user_message; --bare opt-in
  via PLATFORMCTL_CLAUDE_BARE=1 (default OFF — bare disables OAuth +
  Keychain reads; verified via canary #2 self-test)
- `codex_cli.py`: --ephemeral flag added
- `zai.py`: default base URL switched from PaaS (per-call billed)
  to Coding (subscription); /tmp credential-leak pattern removed
  from docstring
- `test_run_review.py`: regression test
  `test_parse_picks_LAST_verdict_block_when_diff_contains_verdict_literal`
  — exercises the bug class (prompt-echo OK + model NOT_OK → must
  return NOT_OK). Would FAIL before parser fix, PASSES after.

## Changed

- `PLATFORM_CHARTER.md` §3 — added "Learning patterns" subsection with
  two charter rules (PR rejection as learning log + Oracle escalation
  trigger discipline). Per Oracle consult.

## Tests

60/60 green (37 in test_run_review.py incl. new regression, 14 consolidator,
4 negative-control, 5 platformctl smoke).

## Linkage

- Closes #6 (rejected per operator architectural decision 2026-05-01)
- Builds on canary #1 findings (state/reviews/PR-5 in PR #6 history)
- Charter rules validated by Oracle session
  `platform-meta-decisions-2026-05-01`

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

pdurlej commented

2026-05-01 17:29:25 +02:00

Author

Owner

Bilateral linkage (per Oracle GPT 5.5 Pro caveat)

This PR is the successor to rejected #6. Per Oracle's "Always link successor PRs both ways" caveat (charter §3 Learning patterns):

Rejected predecessor: #6 (commit history preserved on codex/review-pipeline/canary-findings-pr6 branch — keeping per Oracle's "Keep rejected branches or at least record exact commit SHAs until the successor merges" caveat)
What this preserves: the actual code fixes from canary #1, #2, #3 cycles
What this drops: machine-local state, credential-extraction docs, review-artifact JSONs, ledger noise (everything 4 reviewers convergent flagged)
For diffing rewrites: see git range-diff main...origin/codex/review-pipeline/canary-findings-pr6 main...HEAD (Oracle recommendation)

This is the first exercise of the reject + improved-from-#N pattern enshrined in PLATFORM_CHARTER §3 (this PR also adds the rule itself — a meta-test of the pattern: would an Oracle-validated, narrow successor PR pass review where the bundled predecessor failed?).

## Bilateral linkage (per Oracle GPT 5.5 Pro caveat) This PR is the successor to **rejected #6**. Per Oracle's "Always link successor PRs both ways" caveat (charter §3 Learning patterns): - **Rejected predecessor:** #6 (commit history preserved on `codex/review-pipeline/canary-findings-pr6` branch — keeping per Oracle's "Keep rejected branches or at least record exact commit SHAs until the successor merges" caveat) - **What this preserves:** the actual code fixes from canary #1, #2, #3 cycles - **What this drops:** machine-local state, credential-extraction docs, review-artifact JSONs, ledger noise (everything 4 reviewers convergent flagged) - **For diffing rewrites:** see `git range-diff main...origin/codex/review-pipeline/canary-findings-pr6 main...HEAD` (Oracle recommendation) This is the **first exercise** of the `reject + improved-from-#N` pattern enshrined in PLATFORM_CHARTER §3 (this PR also adds the rule itself — a meta-test of the pattern: would an Oracle-validated, narrow successor PR pass review where the bundled predecessor failed?).

pdurlej added 1 commit

2026-05-01 17:29:26 +02:00

docs(charter): replace codex-fallback wording with Oracle GPT 5.5 Pro actual 13346f9862

Operator pasted the full Oracle response from session
`platform-meta-decisions-2026-05-01`. This commit replaces my
codex-fallback summary in §3 "Learning patterns" with Oracle's actual
wording, which is more complete:

- Adds bilateral linking requirement (operationally critical — Oracle
  caveat #2: "Always link successor PRs both ways")
- Adds `git range-diff` recommendation for comparing rewrites (with
  link to git docs)
- Adds Forgejo PR-as-attempt-history alignment (with link to Forgejo
  docs)
- Adds operational nuance "scope, trust boundary, or review context
  is contaminated" — sharper trigger than my fallback
- Adds 3 stricter Oracle-escalation caveats: don't replace operator
  judgment, ask for decision framing not permission; require
  compressed packet input; convert direction into one operational
  rule then return to deterministic flow

External-source links:
- forgejo.org/docs/next/user/pull-requests-and-git-flow/
- git-scm.com/docs/git-range-diff

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

pdurlej merged commit 0f9ec45b57 into main

2026-05-01 17:39:52 +02:00

pdurlej referenced this pull request from a commit

2026-05-01 17:39:53 +02:00

Merge pull request 'fix(review-pipeline): improved-from-#6 — narrow code-only successor' (#7) from codex/review-pipeline/improved-from-pr6 into main

pdurlej referenced this pull request

2026-05-01 19:20:00 +02:00

feat(identity-wiring): Order B — glm comment hook + charter §3/§4 + identity doctor stub #8