Bundle Algorithm Context

Bundle Algorithm Context

Build the smallest evidence package that can support the decision. Do not dump the repository.

Before writing bridge state, read the canonical protocol at ../gpt-pro-question-window/references/bridge_protocol.md. Before validating a generated package, read references/bundle_schema.md.

Required flow

1. Resolve the external route first with ../gpt-pro-project-workspace/scripts/resolve_bridge_route.py and pass the Codex-decided --external-reasoning or --local-only intent. Use its bridge-thread-id and optional bridge-project-id; do not invent a second identity during bundling.
2. Choose a review mode: algorithm_review, experiment_analysis, paper_brainstorm, implementation_check, or general_question.
3. Inspect repository status, changed and untracked files, relevant architecture/docs, and the user's explicit focus paths.
4. Create an immutable Codex notes snapshot with scripts/prepare_codex_session_notes.py. In Project mode, pass --bridge-project-id.
5. Choose repository context:
   • auto for the first evidence-heavy round. Pass important --include paths as required focus seeds; the builder closes definitely-local dependencies before adding breadth.
   • explicit for follow-ups; name every required file. Safe requested files such as .mjs must appear in the manifest or the build fails.
   • none for reasoning-only follow-ups; --max-files 0 is valid.
6. Run scripts/build_algorithm_bundle.py, normally with --format zip, and pass the same Project ID when applicable. The builder adds a compact, generated Project context file; it does not turn the bundle into a Project Source.
7. Open the manifest and verify every supplied and omitted input, selection reason, dependency closure, repository label, mode-specific output contract, context budget, and safety warning.
8. Upload only after the manifest matches the intended evidence scope.

Completion criterion: the artifact is new and immutable, its zip integrity check passes, no absolute local repository path appears in the manifest, every requested include is present or the build failed, the auto dependency closure is complete, and GPT Pro can distinguish supplied evidence from missing evidence.

Selection rules

Prioritize the problem statement, current notes, changed implementation, configs, data/label construction, training or inference logic, evaluation, metrics, result summaries, and exact files named by the user.

Exclude env files, credentials, cookies, keys, databases, raw private data, vendor trees, generated artifacts, large binaries, and unrelated files. Do not rewrite ordinary source contents. Fail on high-confidence secret-like content unless the exact files have been manually reviewed and explicitly allowed.

Reject missing, filtered, or over-budget explicit includes and paths outside the repository by default. Use --allow-incomplete-includes only after inspecting and accepting every recorded evidence gap. Use --allow-external-include only after confirming every external file; archive names remain anonymized.

Do not use file count as a completeness signal. Selection priority is: explicit focus, question-named definitions, definitely-local dependencies, matching tests and boundary partners, runtime/compiled counterparts, relevant changed files, then general repository context. If a required closure exceeds the budget, increase the reviewed budget or narrow the focus; do not upload a silently truncated closure. Use --allow-incomplete-auto-context only after inspecting every unresolved local dependency and accepting the evidence gap.

Multi-round policy

Do not resend code on every follow-up. Send current Codex notes and compact thread events, adding repository files only when they changed or GPT Pro needs to inspect them again.

The local JSONL ledger remains complete. A bundle includes the latest 24 events and at most 20,000 thread-context characters by default. Bundle construction is not a timeline event; the artifact is recorded only when save_bridge_turn.py captures the exchange that actually used it.

Project Sources and Task Bundles have different lifetimes. Synchronize stable, cross-conversation material through $gpt-pro-project-workspace; keep transient code, diffs, logs, and round-specific questions inside immutable Task Bundles.