entire-context

# Entire context recall (body-free intake + private native tools) **Bounded intake:** On a non-trivial task, the accountable root runs local `status` and one task-scoped body-free `search` before prioritization or dispatch. Consume only verified results relevant to the current request. When private recall would resolve a concrete continuity gap, run the private-mode preflight; a green preflight permits one bounded, repository-scoped native search. It does not require opening every result. Inspect a full checkpoint only when a specific relevant locator materially restores the missing continuity. Do not expand into unrelated tasks, linked sessions, or every private pointer because the tools return them. If the current task is already clear, stop after the bounded intake and continue the work. If relevant local results matter, the root may share one verified body-free `handoff` capsule (at most five cards and 8 KiB) with participants. Children never receive native session bodies. Run `record-use` only when verified local locators materially informed the work. Skip intake for pure one-line questions or a missing/disabled module (note that once); unavailable recall never blocks the canonical task. This skill is the canonical agent-facing contract for the public context layer. It runs one provider-neutral CLI — `python -m scripts.entire_context` — over the local context-link projection. It is **supplemental**: it never replaces GitHub, Fleet Comms, Monitor, session streams, rollover receipts, or formal review, and it never mutates them. ## Hard privacy and scope rules - **Shared capsules are body-free only.** Results distributed to other seats are locator cards and verified canonical excerpts (commit parents/touched paths/timestamps; ACP terminal metadata with `content_included: false`). Never request, emit, or persist transcripts, prompts, responses, subjects, artifacts, raw captures, secrets, or AI-generated summaries. - **Native context stays private to the root.** Never place native prompts, transcripts, responses, session bodies, generated recaps, or generated summaries in a distributed capsule or public evidence. The root may consume search and task-relevant full explain privately after a green preflight. - **No load-bearing Entire dependency.** Entire CLI 0.8.42 stays pinned, optional, and fail-open. A preflight or provider failure falls back to the local body-free workflow and cannot alter task disposition. Native tools never create public `refs/entire/*` or become fleet evidence. - **Fail closed.** Missing, stale, tombstoned, partial-terminal, unsupported, or digest-mismatched evidence is omitted with a machine reason. Never inject an omitted item into your context, and never fabricate coverage for a kind the resolver reports as `unsupported_kind`. - **Query hygiene.** Query text is a ranking needle only: at most 256 UTF-8 bytes, never persisted, never echoed into results. ## Commands Recall commands are local and read-only. Explicit `bootstrap-*`, `reconcile-acp`, and `record-use` commands write only to the rebuildable local projection; provider refreshes write only sanitized local caches. ```bash # Projection state (body-free aggregate) .venv/bin/python -m scripts.entire_context status # Explicit bootstrap/index of real public evidence (idempotent) .venv/bin/python -m scripts.entire_context bootstrap-git <40-hex-sha> [--repo PATH] [--namespace NS] .venv/bin/python -m scripts.entire_context bootstrap-acp conversation_<32hex> [--git-sha SHA] [--acp-root PATH] .venv/bin/python -m scripts.entire_context bootstrap-rollover --agent <agent> --lineage-id <lineage> --rollover-id <rollover> [--rollover-root PATH] .venv/bin/python -m scripts.entire_context bootstrap-github-issue <number> [--issue-cache PATH] [--repo PATH] .venv/bin/python -m scripts.entire_context bootstrap-github-pr <number> --head-sha <40-hex-sha> --repository <owner/repo> [--fleet-root PATH] [--repo PATH] .venv/bin/python -m scripts.entire_context bootstrap-formal-review <review-id> [--fleet-root PATH] .venv/bin/python -m scripts.entire_context bootstrap-fleet-receipt <request-id> [--fleet-root PATH] .venv/bin/python -m scripts.entire_context bootstrap-monitor-run <lease-token> [--monitor-root PATH] # Recover terminal ACP receipts missed by the automatic post-commit callback .venv/bin/python -m scripts.entire_context reconcile-acp --acp-root PATH # search-past-work: ranked verified locator cards (<= 10 results, <= 500 scanned) .venv/bin/python -m scripts.entire_context search --query "<needle>" [--repo PATH] [--acp-root PATH] # explain-change: typed provenance traversal from an exact seed .venv/bin/python -m scripts.entire_context explain-change --sha <40-hex> .venv/bin/python -m scripts.entire_context explain-change --locator-id clink_<64hex> .venv/bin/python -m scripts.entire_context explain-change --canonical-id <id> # prepare-handoff: bounded capsule of verified locators/excerpts (<= 5 items, <= 8 KiB) .venv/bin/python -m scripts.entire_context handoff --locator-id clink_<64hex> [--locator-id ...] .venv/bin/python -m scripts.entire_context handoff --query "<needle>" # After verified locators actually informed the task, attest that use explicitly .venv/bin/python -m scripts.entire_context record-use --task-id <task> --consumer <harness> --purpose <category> --locator-id clink_<64hex> [--locator-id ...] # Explicit Entire 0.8.42 probe for Monitor's sanitized local cache .venv/bin/python -m scripts.entire_context refresh-provider-status [--repo PATH] # Explicit cloud capability probe after private preflight; query/provider bodies are never cached .venv/bin/python -m scripts.entire_context refresh-provider-capabilities --query "<non-sensitive checkpoint canary>" [--repo PATH] ``` The capability probe is operator/root initiated, never part of automatic intake. It repeats the private-boundary preflight, runs repository-scoped search, and sends a dispatch request in the ready source mirror's jurisdiction. It immediately deletes the temporary job; cleanup failure reports a provider error. It persists only booleans, reason codes, HTTP status, and aggregate counts. Monitor reads that cache without a synchronous Entire call. An empty index or dispatch routing failure is a degraded provider feature, not a failure of local body-free recall or of the canonical workflow. Resolution flags: `--repo` (default: cwd) supplies the local git repository; `--acp-root` or `ENTIRE_CONTEXT_ACP_ROOT` supplies the ACP receipt plane root; `--rollover-root` or `ENTIRE_CONTEXT_ROLLOVER_ROOT` supplies the rollover registry state root. Without an ACP root, ACP links fail closed as `source_missing`; without a rollover root, rollover links fail closed as `source_missing`. `--fleet-root`, `--monitor-root`, and `--issue-cache` select existing local canonical stores for fixtures or operators; otherwise linked worktrees use the primary shared `batch_state` stores. `--db` or `ENTIRE_CONTEXT_DB` overrides the projection path. By default every linked worktree resolves the same primary-checkout `batch_state/entire-context/v1` projection through Git's common directory. `--consumer <label>` may be passed by any harness; it is validated, never persisted, and never echoed, so all harnesses receive byte-identical results for identical invocations. An ACP `--git-sha` join is admitted only when hashing that exact SHA matches the conversation's canonical correlation digest. Caller-asserted joins fail closed as `digest_mismatch`. ## Typed resolvers in this slice | Kind | Resolver | Verification | | --- | --- | --- | | `git_commit` | Real | Read-only local git plumbing: parents, touched paths, committer timestamp, author. No commit subject/body. | | `acp_conversation` | Real | Existing terminal receipt verifier (`verify_discussion_receipt`), metadata-only, `content_included: false`. | | `rollover` | Real | Existing read-only registry verifier (`rollover_registry.load_record`): strict body-free projection of schema/key, lifecycle state/boundary, sub-lifecycle states, `cleanup_authorized`, timestamps, and non-body routing (stream epic, issue number, lifecycle state). | | `github_issue` | Real | Fresh `issue_stream_audit.json`: exact open issue plus one unique stream/epic membership. Cache freshness is verification evidence, never locator identity. No body or title. | | `github_pr` | Real | Completed local formal-review job, matching durable `github_publications` row for the exact head and gate context, plus a local Git commit object. | | `formal_review` | Real | Completed job; read-only hash-verified sealed verdict strictly parsed and bound again to its job. Exposes only review/repository/PR/head/gate/state/verdict/model/family/harness/attempt/publication state. | | `fleet_receipt` | Real | Exact terminal Fleet `requests` row only; no invocation specification or messages. | | `monitor_run` | Real | Exact terminal Agent Process Monitor lease only; no PID or process-create time. | All resolver source reads are local and read-only. SQLite opens with URI `mode=ro`; resolvers never migrate, change WAL mode, prune, start a service, or call GitHub, Fleet, Monitor, Entire, or the network. Missing, malformed, nonterminal, stale, unpublished, hash-mismatched, or digest-drifted inputs are omitted fail-closed. ## Private native recall (conditional) For a concrete continuity gap that body-free locators cannot resolve, the accountable root reads [private native recall](references/private-native.md) before the private-mode preflight or any native command. This reference retains allowed command shapes, explicit-request boundaries, privacy restrictions, and failure behavior. It is unnecessary for routine body-free intake. Children never receive native session bodies. Run commands from the repository root. ## How to consume results 1. The accountable root owns recall for the task: run one status/search intake, select only materially relevant verified cards, create no more than one bounded capsule, and distribute it to the relevant participants. 2. Prefer exact identifiers when you have them (commit SHA, conversation ID, locator ID) — exact canonical ID or SHA matches rank first. 3. Treat each relevant card as a **locator**. Read its canonical source (`git show`, the ACP receipt, the GitHub issue) before relying on details. Do not open irrelevant cards or recursively follow every pointer. The `canonical_digest` proves the selected locator matches canonical evidence at recall time; it does not establish relevance to the task. 4. Check `omitted` before concluding absence: an omitted card names the locator and a machine reason (`source_missing`, `digest_mismatch`, `partial_terminal`, `tombstoned`, `unsupported_kind`, `capsule_budget`). 5. A handoff capsule with `"complete": false` intentionally dropped items to stay within the item/byte caps; it is still valid JSON and safe to pass on. 6. Ranking is deterministic and Unicode-casefold based with `locator_id` as the final tie-break, so identical fixtures give identical results to every harness. 7. Search delivery is not evidence of use. When verified cards materially informed intake, architecture, implementation, explanation, review, or handoff, run `record-use` with the exact task, harness, purpose, and locator IDs. The body-free idempotent receipt is the only basis for Monitor's `use.proven` state. ## Harness semantics - Entire integrates with the **host harness**, not each model label. Kimi or GLM running inside Claude Code use the installed `claude-code` integration; those models running inside OpenCode use the installed `opencode` integration. - Codex CLI and Codex Desktop first use the installed native `codex` integration and the same project hooks. Do not invent a `codex-gui` agent. - Claude models hosted by Claude Code use `claude-code`; any supported model hosted by OpenCode uses `opencode`. Grok and AGY follow the integration of the harness that actually launches them. A provider/model label does not identify a capture integration. - Add an external `entire-agent-<harness>` adapter only after a source-blind canary proves that the actual unsupported harness has no native capture path. A model name alone is never evidence that an adapter is required. ## Failure posture A missing, disabled, or unreadable projection yields a body-free status payload (`"available": false`) with exit code 0 for read commands — recall is optional and must never block or mutate your canonical workflow. If recall is unavailable, continue with the authoritative systems directly. A failed private-mode preflight has the same posture: report its body-free issue code, fall back to local body-free recall, and continue. Never weaken the private destination, public-ref, authentication, mirror ACL/readiness, or pinned-version checks to make a native command run.