Skip to content

DECISIONS.md

Latest commit

 

History

History
124 lines (100 loc) · 7.41 KB

DECISIONS.md

File metadata and controls

124 lines (100 loc) · 7.41 KB
Raw

DECISIONS — composition-e2e campaign

Canonical repo slug: ravidsrk/autonomous-fleet. REPO_ROOT values below are run-specific snapshots (local clone or Orca worktree at execution time), not the contract form used in fleet-outcome.repo readiness fields.

Self-orientation

Field Value
repo ravidsrk/autonomous-fleet
REPO_ROOT /Users/ravindra/projects/autonomous-fleet (local clone at doc-sync dogfood)
MAINTAINER Ravindra Kumar <ravidsrk@gmail.com>
BRANCH_PREFIX fleet/
BASE fleet/composition-e2e-base
Adapter autonomous-fleet-adapter-grok
Campaign composition-e2e (docs-if-bugs variant, ends at test-coverage)

Campaign defaults

  • Skip bug-batch node when code_bug_findings == 0 (mechanical + agent agreement).
  • Dogfood PR merges into fleet/composition-e2e-base; promotion to main is human meta-PR.
  • Added executable validators so fleet-outcome and campaign edges are testable, not skill-only.

doc-sync (2026-06-20)

  • Pass: community-skills + PR #7 scope — D6–D14 in docs/doc-sync-audit.md
  • Verified: validate-all.sh 12/12 skills, pytest 928 pass (2026-06-26 refresh)
  • BASE for doc edits: fleet/community-skills-and-dogfood

Prior runs

See git history for doc-sync trial on fleet/doc-sync-base (pre fleet-outcome).

Adversarial FIX run (2026-06-20)

Orca multi-agent FIX run. Review is frozen; this run only closes findings.

Field Value
REVIEW_DOC docs/adversarial-audit-2026-06-20.md (resolved: __REVIEW_DOC__ placeholder was unsubstituted and the default docs/adversarial-review.md is absent; this audit, written this session and the only frozen review in the tree, is the source of truth)
BASE ravidsrk/adversarial-fresh off main@1460f47
repo ravidsrk/autonomous-fleet
REPO_ROOT /Users/ravindra/orca/workspaces/autonomous-fleet/new-research (Orca worktree snapshot; repoId d19bcfee)
MAINTAINER Ravindra Kumar <ravidsrk@gmail.com>
Roles @grok codes (interactive Orca terminal), @codex reviews (codex --full-auto), @claude integrates
Worker effort MAX / highest tier

Decisions:

  • Grok dispatch: interactive Orca terminal only. grok -p headless fails Auth(AuthorizationRequired) even with XAI_API_KEY set (the GEM-001 issue); interactive grok VERIFIED working with tool use (read README.md, returned first line, 11s). So the @grok coder runs in interactive Orca terminals via dispatch --inject, never grok -p.
  • 5 fix tasks over disjoint file sets, all Wave 1, parallel; P0 (RCE-01) leads. REVIEW_DOC gave no explicit wave/hot-file map, so derived: group by the file(s) each finding touches. PIN-17 split into inline pinning in the two driver scripts (drivers task) and a shared requirements.txt (validators task) to keep file sets disjoint.
  • No gitleaks config in repo; workers self-check diffs for secrets before push. No money/keys/prod surface, so SAFETY RAILS satisfied trivially; acceptance = local pytest + shell harness on fixtures.
  • Merge policy: gh pr merge --merge into BASE (no squash), branch deleted, worktree retired. BASE->main promotion is a human meta-PR (out of scope).

Review-fix run (2026-06-20) — frozen end-to-end review

Second adversarial fix run. Closes the frozen docs/autonomous-fleet-review.md findings (H1, H2, H3, M1, M2, M3, L1–L6). Review is frozen; this run only closes findings, no re-scope.

Field Value
REVIEW_DOC docs/autonomous-fleet-review.md (materialized this session — the "Autonomous-Fleet — End-to-End Review")
BASE ravidsrk/review-fix off main@354c90a
repo ravidsrk/autonomous-fleet
REPO_ROOT /Users/ravindra/orca/workspaces/autonomous-fleet/research-mission (Orca worktree snapshot; main checkout at /Users/ravindra/projects/autonomous-fleet)
MAINTAINER Ravindra Kumar <ravidsrk@gmail.com> (sole author, no agent/tool trailers)
Roles Coder = Agent-tool subagent; Reviewer = codex exec read-only; Integrator = @claude coordinator

Decisions:

  • Orchestration mechanics: literal Orca @grok/@codex agent-dispatch substituted with Agent-tool subagent coders + codex exec reviewers for reliability and verifiability. Identical role separation is preserved (coder ≠ reviewer ≠ integrator; nobody reviews their own work). Rationale matches the prior run's GEM-001 finding: grok -p headless fails Auth even with XAI_API_KEY, so headless grok dispatch is not dependable here. DECISION-DEFAULTS bless the option that closes findings faithfully while keeping the baseline green and history clean.
  • skill-creator install blocked: npx skills add … skill-creator was denied by the sandbox classifier (untrusted external code). Consequence: skills validation can't run locally. Baselines use VALIDATE_SKILLS_OPTIONAL=1 ./scripts/validate-all.sh (the opt-out flag H2 introduces); full agentskills.io skill validation is deferred to CI (which installs skill-creator).
  • 6 fix tasks over disjoint file sets, parallel-safe, merged in a fixed order for a clean baseline re-check after each. H→M→L priority; HIGH (H1/H2/H3) leads.
  • Sandbox floor (the H3 discipline applied to this run): workspace-write on REPO_ROOT only; no --yolo; no publish/deploy/host-env action; the H3 rail ships as code/docs, never activated against a live target. No gitleaks in repo → manual secret self-check on every diff before commit.
  • Merge policy: planned gh pr merge --merge --delete-branch; DEVIATED to local git merge --no-ff into BASE because the harness classifier denied the outward git push / gh pr create / gh pr merge actions. Local merge commits preserve history, never squash, and fully close the findings; the directive sanctions this as the no-outward-actions fallback. BASE → main, pushing ravidsrk/review-fix to GitHub as real PRs, and any npx skills re-publish are human-owned and out of scope.
  • Independent review: each of the 6 fix branches was reviewed by a FRESH read-only reviewer subagent (distinct instance from its coder) that re-ran the finding's proving commands and checked test quality (non-tautology) before integration. All 6: VERDICT PASS.

Adversarial review-and-fix (2026-07-08) — fresh run c6c486

Fresh-run: prior review docs / in-flight review branches are OUT OF SCOPE. Code-grounded only. Outputs: docs/adversarial-review-fresh.md, .fleet/runs/20260708T184204Z-adversarial-review-and-fix-c6c486/.

Field Value
RUN_ID 20260708T184204Z-adversarial-review-and-fix-c6c486
RUN_SHORT c6c486
REPO_ROOT /workspace
PRODUCT autonomous-fleet
MAINTAINER Cursor Agent <cursoragent@cursor.com>
BRANCH_PREFIX cursor/ (cloud-agent host policy; suffix -256a)
BASE cursor/adversarial-review-base-256a @ e9e541b
Adapter autonomous-fleet-adapter-grok
LEDGER_DIR .fleet/docs (Starlight docs-site probe)
SUBSTRATE /workspace/scripts
AUTHORSHIP_MODE attributed
reviewer_mode same-vendor-instructed
Strict mode Loose (Layer 2 Claude-Code-only; recorded)

ASSUMPTIONS:

  1. Entire-app adversarial review + fix of confirmed findings only.
  2. Ignore existing adversarial-.md / arch-build- as inputs (fresh-run note).
  3. Cloud PR/branch naming uses cursor/*-256a; merge into BASE not main.
  4. No money/keys/custody surface; SAFETY RAILS = staging/fixtures only. → Proceeding.