docs: clarify agent completion workflow

Tighten the completion checklist so future agents can follow a deterministic start procedure, implementation/doc-tick commit loop, and safe parallelization model. Include one-line prompts for up to three parallel streams with explicit ownership, gates, push, and audit-update requirements.

Co-Authored-By: GPT-5 Codex <noreply@openai.com>
This commit is contained in:
Saravana Achu Mac 2026-05-04 15:55:13 -07:00
parent 381c908ef3
commit 9362f2b843

View File

@ -28,6 +28,133 @@ means, and which gates must pass before pushing.
or auth behavior, pause and confirm unless the audit item explicitly requires
that contract change.
## Fresh-Agent Start Procedure
Every agent should do this before writing code:
1. Read `docs/CODEX_RESUME_PROMPT.md` in full.
2. Read `docs/AUDIT_REDESIGN.md` and this checklist.
3. If touching platform, also read
`/Users/saravana/BytelystAI/learning_ai/learning_ai_common_plat/docs/HANDOVER.md`
and
`/Users/saravana/BytelystAI/learning_ai/learning_ai_common_plat/docs/AUDIT_PLATFORM.md`.
4. Run `source ~/.zshrc && export GITEA_NPM_TOKEN && git status --short`.
5. Confirm the target audit row is still open and no other agent has claimed or
pushed it.
6. Run the relevant baseline gates before making code changes when the work is
not docs-only.
7. Implement only the assigned item or bucket.
8. Run the full relevant gates again.
9. Commit the implementation first.
10. Update the audit row status and fix commit with the short implementation
hash, then commit the doc tick separately.
11. Push both commits.
12. Report the item id, implementation commit hash, doc tick commit hash, and
gate output in three sentences or fewer.
## Exact Commit and Audit Update Loop
Use this loop for every trading audit item:
```bash
source ~/.zshrc
export GITEA_NPM_TOKEN
git status --short
# Baseline gates before substantial code changes
npm --prefix web run typecheck
npm --prefix web run build
npm --prefix web test
npx tsc --noEmit -p backend/tsconfig.json
# Edit files, then repeat gates
npm --prefix web run typecheck
npm --prefix web run build
npm --prefix web test
npx tsc --noEmit -p backend/tsconfig.json
git add <implementation-files>
git commit -m "fix(B2,B3): short description" \
-m "Present-tense why paragraph." \
-m "Refs: docs/AUDIT_REDESIGN.md items B2 and B3." \
-m "Co-Authored-By: GPT-5 Codex <noreply@openai.com>"
IMPLEMENTATION_HASH="$(git rev-parse --short HEAD)"
# Edit docs/AUDIT_REDESIGN.md: ⬜ -> ✅ and Fix commit -> $IMPLEMENTATION_HASH
git add docs/AUDIT_REDESIGN.md
git commit -m "docs(B2,B3): tick ticker header audit rows" \
-m "Record the implementation commit in the redesign audit table." \
-m "Refs: docs/AUDIT_REDESIGN.md items B2 and B3." \
-m "Co-Authored-By: GPT-5 Codex <noreply@openai.com>"
# Final gates after both commits
npm --prefix web run typecheck
npm --prefix web run build
npm --prefix web test
npx tsc --noEmit -p backend/tsconfig.json
git push
```
Do not amend the implementation commit to include the audit-row hash after the
hash is written into the table. That creates a self-referential hash loop. The
preferred pattern is one implementation commit plus one small docs tick commit.
For docs-only work, use the same commit style but run at least
`git diff --check` and the trading gates unless the human explicitly waives
them.
## Parallelization Rules
Maximum parallel streams: 3 agents.
- Parallel streams must own disjoint files or clearly disjoint repos.
- Only one agent may edit `docs/AUDIT_REDESIGN.md` at a time. If multiple
agents finish together, each should pull/rebase first, then update only its
own rows.
- Do not parallelize two items that both touch `web/src/views/HomeView.tsx`,
shared strategy-builder files, or backend route helpers unless one agent is
explicitly acting as integrator.
- Prefer one frontend feature stream, one backend/security/test stream, and one
platform-lint stream.
- Each agent must push incremental commits for its owned item before taking the
next item.
- If a gate fails in code outside the agent's touched files, stop and report
instead of masking the failure.
- If merge conflicts appear in audit docs only, resolve by preserving all
completed rows and commit hashes.
Recommended three-way split for the current backlog:
| Stream | Owns | Avoids |
| --- | --- | --- |
| Agent 1 - Web feature | `B2+B3`, then `B4+B5`, then `B6-B10` | Backend route/cache files and platform repo |
| Agent 2 - Backend/security/tests | `C2`, `F6`, `C1`, then `C3-C7` | `TickerHeader`, layout polish, platform repo |
| Agent 3 - Platform/infra/docs | Platform `P-sweep`, then `E1/E3/E4` if platform is blocked | nomgap WIP files and active trading feature files |
If only two agents are available, run Agent 1 and Agent 2. If only one agent is
available, follow the recommended next commits top-down.
## One-Line Parallel Agent Prompts
Agent 1:
```text
In /Users/saravana/BytelystAI/trading/learning_ai_invt_trdg, read docs/CODEX_RESUME_PROMPT.md, docs/AUDIT_REDESIGN.md, and docs/COMPLETION_CHECKLIST.md, then implement B2+B3 only: wire TickerHeader Watchlist/alert actions and company name from fetchResearchProfile; run all trading gates, commit implementation, tick only B2/B3 rows with the implementation hash in a docs commit, push, and report hashes plus gate output.
```
Agent 2:
```text
In /Users/saravana/BytelystAI/trading/learning_ai_invt_trdg, read docs/CODEX_RESUME_PROMPT.md, docs/AUDIT_REDESIGN.md, and docs/COMPLETION_CHECKLIST.md, then implement C2 only: add a 30-minute in-memory FMP cache keyed by full upstream URL with tests where practical; run all trading gates, commit implementation, tick only C2 with the implementation hash in a docs commit, push, and report hashes plus gate output.
```
Agent 3:
```text
In /Users/saravana/BytelystAI/learning_ai/learning_ai_common_plat, read docs/HANDOVER.md and docs/AUDIT_PLATFORM.md plus the trading docs/COMPLETION_CHECKLIST.md, then do one package-scoped P-sweep lint fix only without touching nomgap WIP or platform pnpm-lock.yaml; run platform typecheck/test/lint, commit that package only, update AUDIT_PLATFORM.md if applicable, push, and report hashes plus before/after lint error count.
```
## Current Snapshot
Completed in trading: