Creatidy/scarcity-router

Fork 0

Correct independent review to use fresh Kilo sessions #215

New issue

Closed

opened 2026-10-23 23:09:04 +00:00 by kilo · 2 comments

kilo commented

2026-10-23 23:09:04 +00:00

Member

Goal

Implement Adrian's 2026-10-23 owner decision that ordinary independent review requires a fresh, separate reviewer conversation, not a separate checkout or execution environment, and that authorized /loop review dispatch/result collection are autonomous through native Kilo session capabilities.

Why

Current repository instructions require the owner to start ordinary reviewer sessions and treat separate-worktree/runtime limitations as blockers. These instructions contradict the owner's intended workflow and prevent autonomous review handoff. Existing Issue #205 / PR #206 is open, review-frozen, and already changes the same broad instruction files; this delivery must remain separate and must not alter its source or consume its review budget.

Scope

Add the authoritative conversational-independence contract to root AGENTS.md.
Reconcile .kilo/command/loop.md, review-pr.md, finish-pr.md, implement-issue.md; .kilo/rules/10-task-system.md, 30-implementation-discipline.md, 35-technical-recovery.md, 40-llm-operating-policy.md, validation.md; and docs/llm-operating-policy.md plus a concise decision-log entry.
Remove the legacy .kilo/agents/pr-reviewer.md; rename the existing standalone reviewer definition to .kilo/agents/pr-reviewer.md and keep it primary-selectable for native Agent Manager sessions.
Apply the owner-specified existing Kilo permission syntax: read-only source/search tools, read-only Forgejo GET/list/search tools, and narrowly allowlisted Git/test commands; deny edit/write/apply_patch/task/external-directory access and shell write/redirection/exec-diff paths. Deny environment files, task-progress files, .git configuration, and local Agent Manager state. The orchestrator supplies the complete diff/source context so routine reads do not prompt the owner.
Update references and the retained deterministic offline reviewer-packet checks/tests for the renamed role and permissions; no live launcher or review service.
Preserve finite review budgets, source freeze/currentness, no self-review or hidden model launcher, and owner-manual Daybreak-through-Codex requirements.

Acceptance Criteria

AGENTS.md states conversational independence is required and filesystem/worktree/container independence is not; same repository, checkout, Git objects, tools and runtime may be reused.
Reviewer receives only issue/task scope, acceptance, exact PR/HEAD references, relevant contracts and complete source diff/instructions; no implementation conversation, reasoning, private deliberations or desired verdict.
Reviewer is a fresh separate session, read-only, and reviews exact frozen source; implementation does not mutate the reviewed source during review.
Authorized /loop autonomously dispatches a visible Kilo reviewer session, collects the native result and proceeds through eligible remediation or subsequent gates without owner prompt/result relay.
/review-pr and /finish-pr describe fresh-session review without requiring the owner to start ordinary Kilo reviews; technical recovery treats a separate worktree as optional absent a demonstrated necessity.
The reviewer profile exposes only the authorized read-only tools, denies writes/delegation/external paths, and keeps sensitive/local state paths denied; complete source is supplied so no routine owner approval is needed.
Review ordinal histories/ceilings, approval/currentness, merge/release/deployment authority, and no-hidden-invocation restrictions remain intact. No formal numbered review is started merely to test setup.
Documentation/workflow checks pass and the final diff contains only this issue's review-workflow instructions, role, and deterministic tests/validator alignment.

Constraints

No product code, Kilo internals, global permission changes, custom orchestrator/wrapper, review service, evidence database, daemon, new permission framework, or changes to frozen PR #206. Do not resume unrelated #204, #206 or #209 work. No review ordinal is reset, transferred, or consumed by setup checks. This issue authorizes only the narrow repository workflow/role correction; ordinary review and integration remain under existing owner-authorized /loop gates.

Links

Related frozen Issue #205: BioMedical-IT/scarcity-router#205
Related frozen PR #206: BioMedical-IT/scarcity-router#206
Owner decision: direct instruction in the 2026-10-23 session.
Relevant docs: AGENTS.md, .kilo/command/{loop,review-pr,finish-pr,implement-issue}.md, .kilo/rules/{10-task-system,30-implementation-discipline,35-technical-recovery,40-llm-operating-policy,validation}.md, .kilo/agents/pr-reviewer.md, docs/llm-operating-policy.md, docs/decisions.md.

## Goal Implement Adrian's 2026-10-23 owner decision that ordinary independent review requires a fresh, separate reviewer conversation, not a separate checkout or execution environment, and that authorized `/loop` review dispatch/result collection are autonomous through native Kilo session capabilities. ## Why Current repository instructions require the owner to start ordinary reviewer sessions and treat separate-worktree/runtime limitations as blockers. These instructions contradict the owner's intended workflow and prevent autonomous review handoff. Existing Issue #205 / PR #206 is open, review-frozen, and already changes the same broad instruction files; this delivery must remain separate and must not alter its source or consume its review budget. ## Scope - Add the authoritative conversational-independence contract to root `AGENTS.md`. - Reconcile `.kilo/command/loop.md`, `review-pr.md`, `finish-pr.md`, `implement-issue.md`; `.kilo/rules/10-task-system.md`, `30-implementation-discipline.md`, `35-technical-recovery.md`, `40-llm-operating-policy.md`, `validation.md`; and `docs/llm-operating-policy.md` plus a concise decision-log entry. - Remove the legacy `.kilo/agents/pr-reviewer.md`; rename the existing standalone reviewer definition to `.kilo/agents/pr-reviewer.md` and keep it primary-selectable for native Agent Manager sessions. - Apply the owner-specified existing Kilo permission syntax: read-only source/search tools, read-only Forgejo GET/list/search tools, and narrowly allowlisted Git/test commands; deny edit/write/apply_patch/task/external-directory access and shell write/redirection/exec-diff paths. Deny environment files, task-progress files, `.git` configuration, and local Agent Manager state. The orchestrator supplies the complete diff/source context so routine reads do not prompt the owner. - Update references and the retained deterministic offline reviewer-packet checks/tests for the renamed role and permissions; no live launcher or review service. - Preserve finite review budgets, source freeze/currentness, no self-review or hidden model launcher, and owner-manual Daybreak-through-Codex requirements. ## Acceptance Criteria 1. `AGENTS.md` states conversational independence is required and filesystem/worktree/container independence is not; same repository, checkout, Git objects, tools and runtime may be reused. 2. Reviewer receives only issue/task scope, acceptance, exact PR/HEAD references, relevant contracts and complete source diff/instructions; no implementation conversation, reasoning, private deliberations or desired verdict. 3. Reviewer is a fresh separate session, read-only, and reviews exact frozen source; implementation does not mutate the reviewed source during review. 4. Authorized `/loop` autonomously dispatches a visible Kilo reviewer session, collects the native result and proceeds through eligible remediation or subsequent gates without owner prompt/result relay. 5. `/review-pr` and `/finish-pr` describe fresh-session review without requiring the owner to start ordinary Kilo reviews; technical recovery treats a separate worktree as optional absent a demonstrated necessity. 6. The reviewer profile exposes only the authorized read-only tools, denies writes/delegation/external paths, and keeps sensitive/local state paths denied; complete source is supplied so no routine owner approval is needed. 7. Review ordinal histories/ceilings, approval/currentness, merge/release/deployment authority, and no-hidden-invocation restrictions remain intact. No formal numbered review is started merely to test setup. 8. Documentation/workflow checks pass and the final diff contains only this issue's review-workflow instructions, role, and deterministic tests/validator alignment. ## Constraints No product code, Kilo internals, global permission changes, custom orchestrator/wrapper, review service, evidence database, daemon, new permission framework, or changes to frozen PR #206. Do not resume unrelated #204, #206 or #209 work. No review ordinal is reset, transferred, or consumed by setup checks. This issue authorizes only the narrow repository workflow/role correction; ordinary review and integration remain under existing owner-authorized `/loop` gates. ## Links - Related frozen Issue #205: https://forgejo.creatidy.com/BioMedical-IT/scarcity-router/issues/205 - Related frozen PR #206: https://forgejo.creatidy.com/BioMedical-IT/scarcity-router/pulls/206 - Owner decision: direct instruction in the 2026-10-23 session. - Relevant docs: `AGENTS.md`, `.kilo/command/{loop,review-pr,finish-pr,implement-issue}.md`, `.kilo/rules/{10-task-system,30-implementation-discipline,35-technical-recovery,40-llm-operating-policy,validation}.md`, `.kilo/agents/pr-reviewer.md`, `docs/llm-operating-policy.md`, `docs/decisions.md`.

kilo commented

2026-10-24 00:57:19 +00:00

Author

Member

Issue #215 is implemented on issue-215-conversational-review; PR #216 is open at HEAD fe470919522e8aada8be3d9ca5d32e9c0e7470c against base cd8fb16b62472b0c4a437a33718c84bbeaaa0ea8.

The worker-anchor validation issue was diagnosed and resolved using a private current-user-owned HOME (0700) with TMPDIR beneath it; no production filesystem security code or test harness was modified. Full make check passes: 2,854 tests, 17 skipped, followed by basedpyright with 0 errors/warnings/notes. Focused workflow/reviewer-packet tests (94) and diff checks also pass.

The PR has undergone fresh native Kilo review sessions; ordinal outcomes and the current #5 reservation are recorded in PR comment #18613. The live native reviewer session/result path has therefore been exercised; no owner prompt/result relay was used. Issue #215 remains open pending authorized integration gates.

Issue #215 is implemented on `issue-215-conversational-review`; PR #216 is open at HEAD `fe470919522e8aada8be3d9ca5d32e9c0e7470c` against base `cd8fb16b62472b0c4a437a33718c84bbeaaa0ea8`. The worker-anchor validation issue was diagnosed and resolved using a private current-user-owned HOME (0700) with TMPDIR beneath it; no production filesystem security code or test harness was modified. Full `make check` passes: 2,854 tests, 17 skipped, followed by basedpyright with 0 errors/warnings/notes. Focused workflow/reviewer-packet tests (94) and diff checks also pass. The PR has undergone fresh native Kilo review sessions; ordinal outcomes and the current #5 reservation are recorded in PR comment #18613. The live native reviewer session/result path has therefore been exercised; no owner prompt/result relay was used. Issue #215 remains open pending authorized integration gates.

adrian.tkacz referenced this issue from a commit

2026-10-24 07:17:44 +00:00

Document conversationally independent Kilo review (#215)

adrian.tkacz referenced this issue from a commit

2026-10-24 07:17:44 +00:00

Align offline reviewer checks with workspace reads (#215)

kilo referenced this issue

2026-10-24 07:17:55 +00:00

Correct independent review to use fresh Kilo sessions #216

~~kilo referenced this issue 2026-10-24 07:18:59 +00:00~~

Correct independent review to use fresh Kilo sessions #216

adrian.tkacz referenced this issue from a commit

2026-10-24 07:38:53 +00:00

Clarify autonomous Kilo reviewer starts (#215)

adrian.tkacz referenced this issue from a commit

2026-10-24 08:44:07 +00:00

Preserve ask-gated reviewer reads (#215)

adrian.tkacz referenced this issue from a commit

2026-10-24 09:49:06 +00:00

Replace standalone reviewer with native Agent Manager role (#215)

adrian.tkacz referenced this issue from a commit

2026-10-24 10:19:57 +00:00

Configure Agent Manager reviewer tool allowlist (#215)

adrian.tkacz referenced this issue from a commit

2026-10-24 10:39:33 +00:00

Enable native Agent Manager reviewer tools (#215)

adrian.tkacz referenced this issue from a commit

2026-10-24 11:39:21 +00:00

Constrain reviewer Git reads to frozen objects (#215)

kilo referenced this issue from a commit

2026-10-24 11:44:50 +00:00

Merge pull request #216: conversationally independent native Kilo review

kilo commented

2026-10-24 11:45:38 +00:00

Author

Member

Completed through authorized canonical integration: PR #216 merged into develop at b2b35bce6678ebc4e393e8c5d78f8c1044f4b119 (2026-10-24T11:44:49Z). Verified merge parents are approved base cd8fb16b62472b0c4a437a33718c84bbeaaa0ea8 and approved HEAD 5f2ffba43e1f75d051b38f675b61f1a35a637197; both are ancestors, and the integrated tree exactly equals the approved/validated HEAD.

Acceptance verified: root conversational-independence contract and D-078; aligned workflow/recovery/validation commands and rules; primary-selectable read-only native reviewer with sensitive-path/Git filesystem restrictions; offline validator/tests only; actual fresh native review dispatch and final-result retrieval with no owner relay. Review #8's P1 was corrected in 5f2ffba; fresh whole-PR review #9 APPROVE with empty findings/limitations is recorded in PR comment #18652, native session ses_e9249bba8ffeqm1sV7jFX8acBj (GPT-6.1 Sol / openai / high dispatch). All #1-9 history is preserved; no #10 dispatch.

Validation: uv sync --only-dev; make check (2,855 tests,17 skipped); uv run basedpyright (0 diagnostics); 95 focused tests; working/staged/full-range whitespace checks PASS. No product/Kilo-internal/global permission changes, wrappers or persistent evidence infrastructure. Frozen PRs #204/#206/#209 and their histories were not modified. Acceptance is satisfied; closing #215, with no release/main/deployment action.

Completed through authorized canonical integration: PR #216 merged into develop at `b2b35bce6678ebc4e393e8c5d78f8c1044f4b119` (2026-10-24T11:44:49Z). Verified merge parents are approved base `cd8fb16b62472b0c4a437a33718c84bbeaaa0ea8` and approved HEAD `5f2ffba43e1f75d051b38f675b61f1a35a637197`; both are ancestors, and the integrated tree exactly equals the approved/validated HEAD. Acceptance verified: root conversational-independence contract and D-078; aligned workflow/recovery/validation commands and rules; primary-selectable read-only native reviewer with sensitive-path/Git filesystem restrictions; offline validator/tests only; actual fresh native review dispatch and final-result retrieval with no owner relay. Review #8's P1 was corrected in `5f2ffba`; fresh whole-PR review #9 APPROVE with empty findings/limitations is recorded in PR comment #18652, native session `ses_e9249bba8ffeqm1sV7jFX8acBj` (GPT-6.1 Sol / openai / high dispatch). All #1-9 history is preserved; no #10 dispatch. Validation: uv sync --only-dev; make check (2,855 tests,17 skipped); uv run basedpyright (0 diagnostics); 95 focused tests; working/staged/full-range whitespace checks PASS. No product/Kilo-internal/global permission changes, wrappers or persistent evidence infrastructure. Frozen PRs #204/#206/#209 and their histories were not modified. Acceptance is satisfied; closing #215, with no release/main/deployment action.

kilo closed this issue

2026-10-24 11:45:44 +00:00