Codebase Memory CLI — Port and Status

This repository is the CLI-only fork and C port of Codebase Memory MCP. The MCP server and third-party MCP integration surface were removed to meet workplace policies that prohibit third-party MCP servers; the local graph, indexing, daemon, watcher, and analysis capabilities remain.

Planning re-baseline (CP61): the CLI-first/MCP-removal architecture phase is closed unless qualification evidence exposes a concrete defect. Remaining work is release qualification: Linux consolidation, immutable GitHub RC artifacts, native Windows validation on the qualification host, frozen-corpus comparative benchmarking, agent usability, production skill, and release decision. See docs/RELEASE_QUALIFICATION_PLAN.md.

Updated: 2026-09-03
Current authoritative baseline: latest complete consolidated source attached to the project sources page
Checkpoint policy: prior overlay/checkpoint archives are historical only and are not replayed or used to reconstruct repository state.
Direction: local CLI code intelligence with protocol-neutral operations and a coordination daemon. MCP is not a supported runtime or integration surface.

Status markers: COMPLETE, PARTIAL, REMAINING, BLOCKED/EXTERNAL.

1. Product / CLI surface

COMPLETE Product executable and user-facing direction are codebase-memory-cli / codebase-memory-cli.exe.
COMPLETE No-argument invocation is CLI help; unknown commands fail as CLI errors rather than starting MCP stdio.
COMPLETE Canonical exploration loop exists for index, projects, status, search, snippet, trace, and coverage.
COMPLETE Neutral graph/source commands now also include schema, query, architecture, changes, source-search, outline, and compare.
COMPLETE Existing cache/index/database naming and persisted compatibility-sensitive internals remain unchanged.
COMPLETE (CP78/CP80 audit): active install, hook-install, update, and uninstall paths do not create, migrate, adopt, or remove codebase-memory-mcp registrations or MCP-owned assets. Historical MCP recognizers may remain as unreachable compatibility utilities, but the CLI lifecycle treats those artifacts as foreign.

2. Neutral operation API / MCP extraction

Neutralized reads

COMPLETE list_projects -> projects
COMPLETE index_status -> status
COMPLETE check_index_coverage -> coverage
COMPLETE search_graph -> search
COMPLETE get_code_snippet -> snippet
COMPLETE trace_path -> trace
COMPLETE (CP36) get_graph_schema -> schema
COMPLETE (CP36) query_graph -> query
COMPLETE (CP37) get_architecture -> architecture
COMPLETE (consolidated CP39 implementation) detect_changes -> changes
COMPLETE (consolidated CP39 implementation) search_code -> source-search
COMPLETE get_file_outline -> outline
COMPLETE compare_graphs -> compare

The authoritative application implementation lives under src/operations/; there is no production MCP compatibility adapter or duplicate MCP handler body.

Protocol-neutral execution support

COMPLETE (consolidated CP39 implementation) bounded shell-command/process-tree execution used by source search and change detection moved out of MCP into src/operations/command_runner.*.
COMPLETE (consolidated CP39 implementation) operation runtime carries cancellation and bounded-command test/limit overrides without an MCP server dependency.
COMPLETE (consolidated CP39 implementation) daemon REQUEST_OPERATION supplies neutral cancellation context to long-running operations.

Remaining read-only audit

COMPLETE: the authoritative MCP-handler audit found no remaining ordinary read-analysis handler after compare_graphs moved to the neutral layer.
REMAINING: continue auditing incidental read helpers only as dependent mutation/admin handlers move; do not create duplicate neutral implementations.

Administrative / mutating operations

COMPLETE: project deletion -> delete-project; authoritative behavior now lives under src/operations/mutation.*, preserves path/alias/tail project resolution through the shared neutral resolver, and refuses uncoordinated execution.
COMPLETE: ADR management -> manage-adr; ADR get/sections/update/set-sections semantics now live under src/operations/adr.*, including legacy-file migration and coordinated writes. The generation-aware store resolver is still supplied through a transitional neutral runtime host seam until store recovery itself leaves MCP.
COMPLETE: trace ingestion -> ingest-traces; current behavior is explicitly non-mutating (counts/accepts supplied observations and reports that runtime edge creation is not yet implemented), so it now lives in the neutral read/administrative operation layer rather than MCP.
COMPLETE: cross-repository intelligence mode now lives in src/operations/cross_repo.*; ordinary index --mode cross-repo-intelligence invokes it directly through the neutral index operation while preserving ordered multi-project leases, wildcard/target validation, cancellation, partial-result semantics, and edge counters.
COMPLETE: ordinary repository indexing now executes through the neutral index operation, including path/project resolution, workspace authorization, pipeline execution, artifact bootstrap, coverage/skip reporting, dump verification, and canonical response construction. Daemon admission/coalescing and supervised-worker containment remain preserved. cross-repo-intelligence now invokes the neutral cross-repository operation directly.

Rule during extraction: preserve behavior first, route all consumers to the neutral implementation, verify parity, delete the legacy body, and only then simplify.

3. Daemon semantics and write coordination

COMPLETE (CP29): distinct neutral REQUEST_OPERATION path exists.
COMPLETE: transitional daemon REQUEST_TOOL vocabulary is retired; all application commands cross IPC through REQUEST_OPERATION.
COMPLETE: explicit daemon REQUEST_MCP vocabulary and the orphaned MCP stdio frontend are removed from production.
COMPLETE/PRESERVED: existing SQLite WAL, busy handling, project mutation leases/locks, worker supervision, staging/atomic publication, cancellation cleanup, and index-job coalescing remain in place.
COMPLETE (consolidated CP39 implementation): long-running neutral reads receive daemon request cancellation without routing through MCP.
COMPLETE for application mutations: neutral mutation operations require explicit runtime authority. Daemon-backed deletion, ADR, indexing, and cross-repository mutation map to the existing cancellable logical reservation plus native per-project lease machinery. Trace ingestion is neutral at its current non-mutating semantics.
COMPLETE: generic tool request vocabulary is retired; neutral operation responses preserve their error bit across IPC without an MCP envelope.
COMPLETE: the generic maintenance observer formerly co-located with the MCP stdio frontend now lives in src/daemon/maintenance_monitor.*; the frontend itself is deleted.
COMPLETE: normal daemon operation/hook sessions use neutral context/cancellation and no daemon application request can instantiate an MCP session object.

4. Runtime Assurance

COMPLETE (CP30): read-only doctor foundation.
COMPLETE (CP30): doctor --deep separates expensive integrity verification from routine checks.
COMPLETE/PRESERVED: SQLite BUSY / LOCKED are treated as transient/busy states, not corruption.
COMPLETE (CP31): stable reliability-event vocabulary and bounded persistent reliability history.
COMPLETE (CP31): doctor can aggregate recent reliability evidence.
PARTIAL: error-code taxonomy and rebuild-reason coverage should continue to be tightened as remaining mutation paths migrate.
REMAINING: benchmark command/baseline comparison surface described in the handoff has not yet been completed.

5. Concurrency / recovery evaluation

COMPLETE (CP32): process-level reader + reader scenario.
COMPLETE (CP32): reader + writer scenario.
COMPLETE (CP32): same-project writer + writer scenario with coalescing/ownership evidence.
COMPLETE (CP32): different-project writer + writer scenario with overlap evidence.
COMPLETE (CP33): controlled writer-client crash/recovery scenario.
COMPLETE (CP33): daemon hard-crash/restart scenario.
COMPLETE (CP34): WAL pressure / long-reader evaluation scaffolding and explicit generation/publication evidence using database identity/generation metadata.
PARTIAL: harness exists, but platform-specific evidence depends on executing a real native Windows binary and representative repositories.

6. Windows 10/11

COMPLETE: native MSYS2/Clang Windows build paths exist; Windows x64 is a release requirement and ARM64 support is present where configured.
COMPLETE (CP35): dedicated Windows validation workflow builds the native codebase-memory-cli.exe and retains evidence artifacts.
COMPLETE (CP35): validation runner includes paths with spaces, Unicode paths, CLI/doctor smoke, concurrency/recovery scenarios, and Windows console cancellation/recovery.
BLOCKED/EXTERNAL: actual Windows validation results must come from the Windows runner/environment; Linux execution is not accepted as Windows proof.
REMAINING: inspect/fix any failures from the real Windows validation run before public release.

7. MCP subsystem retirement

COMPLETE (CP40/CP41): src/mcp/, protocol-only MCP tests, MCP stdio fuzz/introspection tooling, Glama MCP-directory packaging, MCP registry metadata, and MCPB bundle publication machinery are removed from active product/release infrastructure.
COMPLETE (CP41): stale test-runner/TSan references to deleted MCP suites are removed, restoring a coherent neutral test graph.
COMPLETE for read-analysis business logic: all ordinary read-heavy handlers identified in the handoff, plus file outline and graph comparison, now live under src/operations/; MCP is only a compatibility adapter for them.
COMPLETE: the generation-aware store recovery/cache host now lives in src/operations/store_host.*, including integrity classification, confirmed-corruption quarantine, legacy renamed-database fallback, request/idle invalidation, and no-store diagnostics.
COMPLETE: daemon REQUEST_OPERATION sessions own/use the neutral store host directly; neutral operation store resolution no longer reaches through cbm_mcp_server_t or public MCP store-adapter APIs.
COMPLETE for production execution: normal CLI, daemon, supervised worker, and UI paths no longer construct or call cbm_mcp_server_t. The UI /rpc compatibility endpoint maps its two allowlisted calls directly to neutral operations and a neutral store host.
COMPLETE: supervised worker -> daemon coordinator result transport now uses a neutral operation-result wire (payload + is_error) rather than an MCP tool-result envelope.
COMPLETE: reusable JSON argument parsing, supervised-result classification, auto-index file-count admission, tool-profile vocabulary, and CLI/client tool metadata/schema ownership have neutral homes outside src/mcp/.
COMPLETE: the physical src/mcp/ subsystem and protocol-only tests/test_mcp.c suite are removed.
COMPLETE: behavioral indexing/grammar/incremental/integration test harnesses no longer construct an MCP server; retained suites use a neutral test operation host over cbm_operation_execute.
COMPLETE: the empty MCP_SRCS Makefile category is removed; production and test source lists no longer require MCP sources.
COMPLETE: reusable JSON argument parsing used by test/daemon surfaces has a neutral src/operations/json_args.* home.
COMPLETE for active invariant/security harnesses: scripts/smoke-invariants.sh now exercises the CLI-first workflow rather than MCP initialize/tools-list protocol behavior, and scripts/security-audit.sh no longer expects src/mcp/mcp.c.
COMPLETE (through CP47): active release/package/smoke execution and distribution identity use codebase-memory-cli; remaining old repository URLs refer to the historical GitHub slug, and negative smoke/security checks intentionally reject retired MCP artifacts.
PARTIAL: MCP-shaped tier-profile renderers and ownership markers remain only where update/uninstall needs byte-accurate recognition of previously released files. They are legacy removal compatibility, not active install output.
COMPLETE (post-CP47 audit): active installer metadata no longer advertises MCP as a client capability; legacy cleanup is modeled explicitly as a removal obligation rather than a product feature.

No production or retained C behavioral-test source includes mcp/mcp.h or calls cbm_mcp_server_* / cbm_mcp_handle_tool. The obsolete protocol implementation is physically absent from the tree.

COMPLETE (CP53): runtime-isolation wiring for CLI security fuzz/network harnesses uses the shared private runtime/cache helper; the stale MCP-fixture fuzz harness is replaced by a CLI fixture that proves isolation and crash rejection.
COMPLETE (CP53): stale transition vocabulary was removed from active worker/hook/pipeline comments and the phantom mcp suite label was removed from the parallel test scheduler.
COMPLETE (CP53): obsolete tests/DEPRECATED_MCP_TESTS.md was deleted; its inventory no longer described the retained neutral test graph.

8. Release readiness

Internal Windows validation build

READY TO RUN: CP35 provides the dedicated native-Windows validation workflow. It should be run while operation extraction continues elsewhere.

Public CLI-first release

NOT READY YET. Primary blockers:
1. Native Windows validation evidence must be reviewed.
2. Final high-value end-to-end/concurrency/recovery/release verification remains.
3. Post-MCP simplification and compatibility classification must finish without weakening ownership-aware legacy cleanup.

9. Consolidated-source policy

COMPLETE: the latest complete project-source archive is the sole implementation baseline.
COMPLETE: prior CP/overlay artifacts are disposable history and must not be replayed to reconstruct source state.
COMPLETE: this status document is reconciled to the actual consolidated tree rather than inferred from old checkpoint labels.

10. Immediate planned sequence

COMPLETE: neutral mutation-authority contract mapped to the daemon's existing logical reservation + native project lease machinery.
COMPLETE: project deletion migrated as the first bounded administrative mutation, including legacy project argument/path/tail compatibility.
COMPLETE: ordinary indexing extraction: neutral ingress, physical pipeline/response implementation, and worker supervisor are outside MCP while daemon coalescing, worker containment, cancellation, staging/publication, and rebuild classification remain preserved.
Reconcile results from native Windows validation; fix platform-specific locking/cancellation/publication defects before release.
COMPLETE: neutralize the generic store recovery/cache host and make daemon operations use it without an MCP session dependency.
COMPLETE: extract neutral session context/cancellation state and make normal daemon operation/hook sessions avoid MCP allocation.
COMPLETE: retire transitional REQUEST_TOOL; CLI/daemon application commands now use only the neutral operation request and preserve operation error status without MCP envelopes.
COMPLETE: split the generic maintenance observer into src/daemon/maintenance_monitor.* and delete the orphaned REQUEST_MCP/JSON-RPC stdio frontend.
COMPLETE for production: remove direct UI MCP dispatch and neutralize production utility/schema/profile ownership; unlink src/mcp/mcp.c from the product build.
COMPLETE for runtime/test architecture: migrate/delete MCP-only tests and physically delete src/mcp/.
COMPLETE through CP47: active installer/release/package/smoke execution is CLI-first; retain only ownership-aware legacy removal state, persisted compatibility paths, negative guards, and historical repository URLs.
IN PROGRESS: post-MCP architectural simplification audit: transition-only wrappers/vocabulary and the Windows MCP-protocol guard dependency are removed; continue reconciling the remaining soak/repro protocol-shaped reliability harnesses while preserving correctness and legacy removal safety.
Complete benchmark/baseline tooling and run the high-value end-to-end/concurrency/recovery/release verification appropriate for the milestone.

11. Drift assessment

On target. Runtime Assurance and concurrency work was front-loaded before the second read-heavy extraction group. Subprocess cancellation/supervision, indexing, ADR, deletion, trace-ingest, and cross-repository behavior have crossed the neutral operation boundary without duplicate authoritative MCP handlers. CP47 is already merged into the authoritative source. The current phase is therefore post-MCP simplification: remove transition-only vocabulary, dead creation helpers, and stale active product identity while preserving legacy owned-state cleanup, persisted locations, and coordination/recovery semantics. An explicit final Linux production link is being re-established from this authoritative post-CP47 baseline before this document claims build verification; native Windows validation remains separate evidence.

12. Post-CP47 simplification audit

COMPLETE: authoritative baseline confirmed to have src/mcp/ physically absent and no production cbm_mcp_server_*, cbm_mcp_handle_tool, REQUEST_MCP, or REQUEST_TOOL references.
COMPLETE: removed three dormant MCP creation/preflight helpers that had no callers and caused the post-CP47 production build to fail its own -Werror=unused-function policy.
COMPLETE: renamed the stale generic daemon MCP_CLIENT process role to BOOTSTRAP_CLIENT; default/unknown public invocations remain stateless and the retained bootstrap path is protocol-neutral.
COMPLETE: removed dead JSON-RPC/MCP bootstrap formatting/watchdog residue from the generic product entrypoint.
COMPLETE: agent-client registry metadata no longer models MCP as an active capability or carries a one-implementation removal callback; profiles record only whether legacy MCP cleanup applies and call the shared removal routine directly.
COMPLETE: removed no-op tier-profile install wrappers/calls. Ownership-aware tier-profile renderers remain reachable only from update/uninstall cleanup so previously released MCP-bound files can be recognized conservatively.
COMPLETE: corrected stale active product identity in workspace guidance, update-manager suggestions, UI process monitoring, graph UI messaging, generated artifact comments, and the native-Windows guard wrapper.
PRESERVED: legacy codebase-memory-mcp cache/config locations, old owned-entry names/markers, cleanup matchers, negative security/release guards, and historical GitHub repository URLs remain where changing them would break persisted compatibility, safe removal, or repository addressing.
COMPLETE: explicit Linux production link succeeds as build/c/codebase-memory-cli; changed production C surfaces compile under the project's -Wall -Wextra -Werror policy.
COMPLETE: removed the obsolete MCP-stdio parent-watchdog harness from the active test graph; retained the supervised worker parent-death watchdog because it protects a real CLI worker containment property.
COMPLETE: active runtime/regression scripts that build or execute the product now default to build/c/codebase-memory-cli; canonical benchmark parsing no longer carries an MCP content-envelope fallback.
COMPLETE: graph-UI missed-coverage guidance now instructs agents to use the canonical CLI status command rather than an MCP tool call.
COMPLETE (CP49): native-Windows product guards no longer depend on the retired MCP stdio/JSON-RPC client. Unicode repo/cache indexing now exercises canonical index/query; daemon churn uses canonical projects; update handoff verifies that a live permanent daemon retains the same pid; the protocol-only tests/windows/mcp_stdio.py helper is deleted.
COMPLETE (CP50): active watcher, memory-lab, and search-performance harnesses no longer construct MCP/JSON-RPC traffic. The watcher kill-switch is exercised through config/daemon/index; memlab repeatedly drives canonical operations against the long-lived coordination daemon; the search benchmark invokes canonical search directly.
PRESERVED: daemon stop refusal while committed clients exist remains covered directly by the neutral daemon runtime suite. The obsolete Windows product-surface setup that created such a client by opening an MCP stdio session was removed rather than replaced with a fake frontend.
BLOCKED/EXTERNAL: native Windows execution/validation evidence.
COMPLETE (CP51): scripts/soak-test.sh and tests/repro/issue832_rss.py no longer drive a retired MCP stdio/JSON-RPC frontend. Soak now keeps the supported permanent coordination daemon under diagnostics, exercises canonical CLI read/index operations, and kills the daemon itself during an active index to validate recovery. The #832 RSS repro samples the permanent daemon across canonical index cycles for in-process versus supervised isolation.
COMPLETE (CP52): tests/test_daemon_smoke.py is now a CLI-first real-binary guard. It covers cold/warm CLI clients, permanent-daemon PID stability, daemon-backed indexing, crash recovery, and clean stop. Detailed cancellation, session isolation, rendezvous/version-conflict, and committed-client stop-refusal semantics remain in the neutral tests/test_daemon_runtime.c suite rather than being duplicated through a retired frontend protocol.

CP54 — Public documentation and release-contract reconciliation

COMPLETE: Active homepage and llms.txt product identity now describe codebase-memory-cli as a local CLI-first code-intelligence application rather than an MCP server. Historical repository/homepage URLs remain unchanged because the repository slug is compatibility/history, not runtime architecture.
COMPLETE: docs/CONFIGURATION.md distinguishes compatibility-sensitive persisted codebase-memory-mcp paths from the active CLI product and removes stale MCP-only request-path wording.
COMPLETE: version- and language-count contract surfaces now point at the renamed CLI package files; the vacuous npm language-count surface was removed from the contract because it no longer publishes a count.
COMPLETE: active test/workflow comments and usage examples touched by this slice no longer describe neutral indexing/runtime behavior as MCP behavior.

CP55 — Internal profile/observability and UI protocol cleanup

COMPLETE: removed the dead cbm_log_mcp_request() observability helper and its self-only unit coverage; production had no callers after physical MCP removal.
COMPLETE: removed the daemon ALL/ANALYSIS/SCOUT tool-profile field and src/operations/tool_profile.h. Every production context caller already used the unrestricted value, so the profile byte was transition-only state rather than a live product choice.
COMPLETE: simplified daemon SET_CONTEXT wire state by removing the unused profile byte while preserving root, allowed-root, hook event/dialect, cancellation, watcher, auto-index, update, and UI coordination behavior.
COMPLETE: migrated retained daemon-application lifecycle tests from retired numeric request kind 2 / JSON-RPC pings to neutral operation requests or direct context establishment; deleted tests whose only contract was MCP notifications/restricted MCP surfaces.
COMPLETE: removed the UI POST /rpc JSON-RPC compatibility endpoint. The current embedded UI has no consumer for it; supported UI behavior remains on neutral /api/... routes and direct neutral operations/store infrastructure.
COMPLETE: migrated retained HTTP security/watchdog tests to supported /api/... endpoints and deleted the RPC allowlist test that protected only the retired protocol envelope.
PRESERVED: tier-profile renderers and exact --tool-profile=analysis|scout strings remain only in ownership-aware uninstall/update recognition of files produced by historical releases. Clean install does not expose a daemon profile mode.
PRESERVED: historical repository URLs, persisted codebase-memory-mcp cache/config paths, owned-entry markers, and negative guards remain compatibility/history rather than active architecture.

CP56 — Legacy profile compatibility boundary and doctor configuration visibility

COMPLETE: renamed the historical tier/profile renderer source and test surfaces to legacy_agent_profiles.*, making their ownership explicit: they reproduce bytes from prior releases solely so update/uninstall can remove exact Codebase Memory-owned documents without deleting modified or foreign files.
PRESERVED: historical --tool-profile=analysis|scout, MCP server names, and old agent dialect payloads remain inside that legacy renderer because byte-accurate recognition is a safety mechanism, not an active install capability.
COMPLETE (CP78): OpenClaw compaction augmentation uses the CLI-owned Codebase Knowledge Graph (codebase-memory-cli) label. The CLI lifecycle no longer claims the historical MCP label.
COMPLETE: the durable agent-instructions contract test now validates the actual CLI-first instructions and rejects codebase-memory-mcp/search_graph leakage from new instructions.
COMPLETE: doctor now reports effective watcher state plus UI enabled/port configuration in both human and JSON output, reusing existing configuration sources rather than introducing duplicate health infrastructure.

CP57 — creation-era test and dead result-helper cleanup

COMPLETE: removed the unused cbm_cli_mcp_result_is_error() API. It had no production caller and survived only through a self-test for the retired MCP result envelope.
COMPLETE: removed creation-era CLI tests that expected fresh tiered MCP subagents, per-agent MCP server blocks, or --tool-profile registrations. Current install behavior is CLI-first; byte-accurate historical renderers remain tested separately for ownership-aware update/uninstall cleanup.
PRESERVED: durable CLI instructions, skills, hooks, foreign-file protection, and historical recognition helpers remain where useful for compatibility analysis; CLI lifecycle operations do not use those helpers to remove MCP-owned state.

CP58 — recovery observability and concurrency evidence audit

COMPLETE: audited the retained recovery/concurrency harnesses and confirmed substantive coverage already exists for cross-process lock ordering/fairness, cancellation rollback, WAL crash recovery, mutation serialization, index coalescing, worker containment, daemon stop/refusal, and corruption quarantine/recovery. No duplicate concurrency framework or broad low-value unit-test expansion was added.
COMPLETE: the neutral store host now records reliability events when a previously suspect persisted generation rechecks as transient, corrupt, successfully quarantined, or requiring rebuild. A successful coordinated recheck is also recorded as integrity-ok.
COMPLETE: human doctor output now surfaces observed bounded reliability-event counts in addition to the existing JSON summary; unobserved/reserved event identifiers remain absent rather than being presented as synthetic zero evidence.
PRESERVED: BUSY/LOCKED classification remains distinct from corruption; the new event wiring records only verdicts actually produced by the existing recovery boundary and does not infer SQLite error codes it does not own.
REMAINING: establish real workflow performance baselines and run the phase-end Linux reliability pass; native Windows validation remains external evidence.

CP59 — Agent-workflow performance harness

PARTIAL — The performance-baseline harness is now CLI-first and reproducible. The canonical workflow runner covers index, search, architecture, snippet, outline, changes, status, and optional cross-repository intelligence using an isolated cache. scripts/benchmark-index.sh no longer uses the deprecated raw cli index_repository compatibility interface.

Numeric baseline capture remains REMAINING for this checkpoint because the production -O2 build did not complete within the available uninterrupted compiler window; no substitute unoptimized/older binary was used for measurements. See docs/PERFORMANCE_BASELINE.md.

CP60 — Benchmark reproducibility contract

COMPLETE: the workflow benchmark now defines a reproducibility contract rather than treating raw timings as self-validating evidence. Comparative runs must pin executable identity/hash, benchmark-repository commit and dirty state, harness revision/parameters, cache-state procedure, host metadata, and relevant CBM_* environment.
COMPLETE: each benchmark invocation is one independent trial and refuses a reused result/cache directory. Repeated read cases receive one unrecorded warm-up before recorded timings so reference/candidate comparisons use the same steady-state semantics.
COMPLETE: environment.txt records binary/harness SHA-256, repository revision/origin/dirty state, timestamp, CPU, memory, filesystem, workload parameters, and benchmark-relevant environment where available.
COMPLETE: the documented release comparison procedure calls for at least three independent fresh-cache trials per binary, median-based read comparison, exact input/harness parity, correctness/graph-shape checks, and explicit invalidation conditions for mismatched runs.
REMAINING: capture numeric reference/candidate baselines with a successfully linked production-optimized binary. The methodology is now explicit enough for those results to be independently repeated; numeric evidence is still not claimed by this checkpoint.

CP61 — Release qualification planning re-baseline

COMPLETE: migration-era sequencing is superseded by docs/RELEASE_QUALIFICATION_PLAN.md; further architecture changes require evidence from qualification rather than generic cleanup.
COMPLETE: native Windows release validation is specified as an external procedure against immutable GitHub-published bytes on the qualification host; portable and installed forms are both required.
COMPLETE: comparative Windows benchmarking is specified around a frozen repository corpus, exact artifact identity, fresh-cache alternating trials, correctness/shape parity, and retained raw evidence.
COMPLETE: a standalone Windows qualification-agent runbook defines acquisition, identity verification, frozen-corpus checks, functional/recovery testing, install/update/uninstall validation, benchmarking, failure classification, and evidence summary requirements.
COMPLETE: benchmark corpus revisions are immutable within a corpus generation. Any repository revision change creates a new generation and requires a new baseline capture rather than comparison against older results.
REMAINING: finalize the exact agent-workflow remote and select/pin the fourth larger open-source repository already available on the CI host / the qualification environment.
REMAINING: complete Linux consolidation, publish an immutable GitHub release candidate, execute native Windows qualification and comparative baselines, then run agent usability and produce the final product skill.

CP62 — Frozen benchmark corpus finalization

COMPLETE (planning): windows-corpus-1 now has fixed repository identities: codebase-memory-cli, specgen-aw, agent-workflow, agent-workflow-spec-contracts, and managed terminal. Repository remotes are explicit in docs/qualification/BENCHMARK_CORPUS.json. Exact commit SHAs remain deliberately unset until the one-time initialization reads them from the actual clean frozen checkouts used on the qualification host; after pinning, any commit change creates a new corpus generation and requires new baseline capture.

CP63 — Linux qualification consolidation

Status: IN PROGRESS / QUALIFICATION

The project is now executing the release-qualification plan rather than further migration cleanup. The fast sanitized foundation gate has been repaired to use a true foundation-only incremental runner and currently passes 312 tests with two explicit Windows-only skips. Release/runtime/security/smoke contracts have been reconciled with the supported CLI-first release topology and pass. See docs/LINUX_RELEASE_QUALIFICATION.md for evidence and remaining Linux phase-end work.

The current optimized production build has not yet reached final link within this execution environment's bounded compiler windows; no production compiler error has been observed. Final-link evidence remains required for the current candidate and will also be obtained from the immutable GitHub RC build.

CP64 — external qualification release hold

COMPLETE: GitHub release preparation now defaults to a draft external-qualification hold. A separate promotion workflow validates the qualification host qualification evidence against the exact Windows release archive/executable hashes before registry publication or public un-drafting. Linux broad sanitized qualification remains in progress due build duration; no source failure has surfaced in the incremental build.

CP65 — RC dispatch reproducibility (release qualification)

Status: COMPLETE

Added a deterministic RC-dispatch preparation tool that refuses dirty trees and non-vX.Y.Z-rc.N tags.
Added a standalone GitHub RC dispatch agent runbook.
RC dispatch evidence now binds source commit, workflow hash, corpus manifest hash, qualification host, and held-draft release inputs.
The first CLI-first RC is recommended as v0.11.0-rc.1 after the current public v0.10.8, but version selection remains an explicit operator decision.
Broad Linux ASan/UBSan runner compilation remains in progress locally; no compiler/source failure has surfaced in the preserved incremental build.

CP66 — explicit first-CLI baseline decision

Status: COMPLETE (release qualification contract)

RC dispatch now refuses to proceed without an explicit benchmark-baseline decision: either --baseline-zero or an immutable --baseline-tag.
The first CLI-first candidate (v0.11.0-rc.1 under the current plan) must explicitly use BASELINE_ZERO; v0.10.8 is not treated as a like-for-like CLI workflow performance baseline.
BASELINE_ZERO is never inferred automatically. After the first qualified CLI release, later candidates must name the exact prior qualified CLI release tag for comparative benchmarking.
The Windows qualification runbook consumes the RC dispatch decision and must not substitute an incompatible baseline artifact.
Local broad sanitized compilation remains useful supporting evidence, but the held GitHub RC workflow is the authoritative final build/test source for immutable release bytes.

CP74 — CP73 qualification-harness rebase onto cf0869fc

Status: IMPLEMENTED / REQUIRES EXTERNAL RERUN

Rebases the CP73 qualification-harness correction onto the newer authoritative cf0869fc source without reconstructing or overwriting unrelated newer work.
Preserves the newer smoke-script revisions already present in cf0869fc; their --help and strict unknown-option contracts already satisfy the CP73 interface requirement and therefore are not part of the CP74 payload.
Restores interpreter-first venue-parity dispatch so Python entry points are invoked only with python3, never first through bash.
Restores the CP73 memory-analysis classification: Heaptrack retention is diagnostic, fork-child leak classification cannot rewrite behavioral child exit codes, and Memcheck error contexts plus top-level definite/indirect ownership loss remain release failures.
Preserves the newer 2026-09-04 full-suite remediation record in docs/LINUX_RELEASE_QUALIFICATION.md and appends the CP72/CP74 qualification evidence instead of replacing it.
Focused interface/venue-parity validation is required for the checkpoint; the canonical full sanitized suite and dynamic memory-analysis leg still require a fresh external rerun before qualification is called green.

CP75 — immutable RC identity enforcement

COMPLETE: the production release workflow no longer exposes replace=true, deletes an existing release/tag, or force-moves a release tag.
COMPLETE: immediately before drafting, the workflow refuses an already-existing GitHub release or remote tag and creates the dispatched-SHA tag with a normal non-force push, so duplicate/racing dispatches fail closed.
COMPLETE: the existing release gate-chain contract now protects this immutable-identity boundary; the RC dispatch runbook records that workflow-side enforcement is independent of the operator's pre-dispatch availability check.
PRESERVED: the external-qualification hold remains the default and registry/public promotion remains owned by the promotion workflow after exact Windows-byte qualification.

CP76 — Post-qualification CLI contract reconciliation

COMPLETE: the first broad sanitized run after CP74 reached suite execution and isolated all 59 failures to cli (53) and incremental (6); the daemon application/runtime/IPC suites were green.
COMPLETE: the six incremental failures were stale assertions against retired MCP-shaped/raw response text. They now assert the neutral operation-result keys/statuses actually emitted by the production operation layer.
COMPLETE: stale CLI installer tests now describe supported CLI-first durable context, skill, config, and lifecycle surfaces instead of requiring fresh MCP registration. Legacy MCP ownership remains tested only where exact cleanup/migration compatibility is still product behavior.
FIXED: registry-backed agent installs now resolve their effective client config path before Qoder/Devin lifecycle installation and cleanup; previously those hooks could be silently skipped as config path unresolved.
FIXED: hook project resolution no longer opens a derived, nonexistent project DB as a fast probe. The derived-name path is used only when that exact DB already exists, allowing indexed custom project names to fall back to canonical root_path metadata instead of being shadowed by a phantom empty DB.
FIXED: exact-owned Claude gate-script migration recognizes the released MCP-era sibling binary basename as well as transition-era branding/path combinations. Ownership remains byte-exact; modified/foreign scripts are still refused.
VERIFIED: the final eight-case CLI sanitizer cluster passes 8/8 under ASan/UBSan with leak detection enabled. The changed production and test objects compile under the repository's strict sanitizer and production warning policies.
PENDING EXTERNAL GATE: rerun scripts/test.sh on the CP76 candidate. CP76 changes production CLI code, so the earlier 7,352-pass/59-fail broad result is diagnostic input, not final Gate-1 evidence. The corrected post-CP74 memory-analysis run is already green and need not be repeated unless the broad sanitizer rerun produces new memory evidence.

CP77 — Incremental Gate-1 closeout

EVIDENCE: the CP76 canonical rerun completed all 138 suites with 7,408 passed / 3 failed / 3 skipped. cli is 236/236 and all daemon suites are green; only incremental remains red at 160/163.
FIXED (test host): the neutral in-process operation host now wires project_invalidate to its cached store host. Reindex tests therefore release ADR/read-side cached handles before physical publication, matching the separate-process lifecycle used by production indexing.
FIXED (read path): search_graph no longer uses a create-capable open for an unknown project. It requires an existing read-only project DB with the requested project row, so an invalid search cannot create a ghost empty database or masquerade as an empty successful result.
RECONCILED: the remaining count-query assertion now checks the current compact table contract (rows: ... (cols: n)) rather than a retired JSON/MCP-shaped key assumption.
MEMORY ATTRIBUTION: the CP76 incremental LeakSanitizer total (2,896 bytes in three allocations) exactly corresponds to response buffers abandoned by those three failed assertions; the independent CP76 memory-analysis leg is green with 0 Valgrind lost bytes.
VERIFIED LOCALLY: changed production/test objects compile under the repository's strict production and ASan/UBSan warning policies. Full incremental execution still requires the externally cached FastAPI fixture unavailable in this sandbox.
PENDING EXTERNAL GATE: run scripts/test.sh once on CP77. No separate memory-analysis rerun is required unless new memory evidence appears.

CP78 — Explicit CLI hook-install boundary and MCP coexistence

COMPLETE (implementation): ordinary codebase-memory-cli install is an asset-only surface. CLI-owned agent hooks require the separate, explicit codebase-memory-cli install-hooks command; --plan, --dry-run, and --clients are independently supported on that hook surface.
COMPLETE: CLI-owned durable-context assets and hook identities use codebase-memory-cli namespaces rather than MCP-era names. The CLI no longer adopts MCP-era Claude/Gemini hook identities as upgrade ownership.
COMPLETE: install, hook installation, and uninstall treat codebase-memory-mcp registrations, MCP-era hook files, legacy MCP compaction labels, and MCP-owned skill namespaces as foreign. Side-by-side installation therefore does not use MCP cleanup as a migration mechanism.
COMPLETE: setup scripts no longer activate repository Git hooks implicitly. Post-install guidance makes agent-hook installation explicitly opt-in.
REGRESSION COVERAGE ADDED: CP78 plan tests assert that asset and hook plans are disjoint, and a Claude coexistence test asserts that explicit CLI hook installation preserves a pre-existing MCP hook command/file while installing separate CLI hook identities.
VERIFIED LOCALLY: changed production CLI and CLI-test translation units compile with the repository's strict -Wall -Wextra -Werror production/test policies and ASan/UBSan test instrumentation.
VERIFIED LOCALLY (runtime): the two CP78-specific sanitizer regressions pass 2/2: asset and hook plans are disjoint, and explicit Claude CLI-hook installation preserves pre-existing MCP hook files/settings while adding separate CLI-owned hook identities. The broader CLI suite cannot be treated as local Gate-1 evidence in this sandbox because an older activation/hash test stalls before the CP78 region when the runner is relocated under a private safe path.
PENDING EXTERNAL GATE: run the canonical Jenkins/scripts/test.sh gate on the CP78 candidate; reconcile any stale pre-CP78 ownership assertions rather than weakening the MCP/CLI separation contract.

CP79 — CP78 local qualification closeout

COMPLETE: CP78's focused runtime qualification is green under ASan/UBSan (2/2). The coexistence regression was corrected to snapshot file contents with allocating reads rather than aliasing the test helper's static buffer.
COMPLETE: a fresh incremental CLI runner was linked successfully after the repository-wide sanitizer dependency graph completed. Running from /mnt/data correctly triggers the installer's unsafe-ancestor guard; relocating the runner beneath a private 0700 directory gets past that guard but an unrelated pre-existing activation/hash test stalls before the CP78 assertions. No product workaround was added.
QUALIFICATION LIMITATION: an optimized production build progressed through all ordinary production objects and grammar objects but the monolithic prod_lsp_all.o compilation alone exceeds the sandbox's maximum uninterrupted command window (including a 240-second attempt). No production compiler error surfaced.
PENDING EXTERNAL GATE: Jenkins remains the authoritative Gate-1 venue for the complete clean sanitizer suite and optimized production link. CP79 makes no runtime/product architecture changes beyond the CP78 candidate.

CP80 — Static side-by-side ownership audit

FIXED: KiloCode asset installation once again installs the CLI-owned codebase-memory-cli.md durable instruction and registers only that instruction reference; the CP78 reconstruction had accidentally left uninstall operations in the install branch.
FIXED: KiloCode uninstall now removes only the CLI-owned instruction reference/file and leaves both current and historical MCP configuration/rules untouched.
FIXED: shell PATH persistence is labeled codebase-memory-cli install rather than the retired MCP product name.
RECONCILED: migration-status language now matches the CP78 side-by-side contract: active CLI lifecycle operations do not use historical MCP cleanup as a migration mechanism.
STATIC AUDIT ONLY: per user direction, CP80 was reviewed by source/diff inspection only. No compilation or tests were run for this checkpoint.

CP81 — Static lifecycle ownership closeout

FIXED: VS Code profile uninstall no longer enumerates profile mcp.json files or removes any MCP registration; profile MCP state is foreign under the CP78 side-by-side contract.
FIXED: hook-augment timeout breadcrumbs now use the CLI-owned ~/.cache/codebase-memory-cli/logs namespace rather than writing diagnostics into the MCP cache namespace. The compatibility-sensitive graph/index cache remains unchanged.
RECONCILED: cli.h, agent-client source comments, docs/llms.txt, configuration guidance, and Windows/Linux qualification instructions now describe the actual split lifecycle: install writes assets only, install-hooks is explicit, and MCP state is preserved.
STATIC AUDIT ONLY: no compilation or tests were run for CP81 by explicit instruction. Package verification is limited to source/diff/overlay integrity checks.

CP82 — Static lifecycle wording and legacy-fingerprint audit

STATIC AUDIT ONLY: no compilation or tests were run for CP82.
RECONCILED: generic asset installation now reports that hooks require the explicit codebase-memory-cli install-hooks surface instead of describing the asset-only path as “instructions/hooks”.
RECONCILED: uninstall --help now explicitly limits removal to CLI-owned agent assets/hooks and states that MCP registrations/assets are preserved.
CLARIFIED: legacy MCP SessionStart/SubagentStart payload constants are named and documented as byte-for-byte recognition fingerprints only; current CLI hook generation continues to invoke codebase-memory-cli hook-augment.
REVIEWED: release-facing codebase-memory-mcp repository URLs remain intentional historical-repository-slug references; compatibility cache/config paths retaining the historical name are likewise unchanged.

CP83 — integration version reconciliation (static)

COMPLETE (static only; no compile/test run): reconciled qualification metadata with the current SpecGen/Agent-Workflow integration set: SpecGen 0.2.1, Agent-Workflow 0.10.0, and released specgen-agent-workflow-contracts 0.2.1. The benchmark corpus now records these as expected release versions while retaining immutable commit SHAs as the authoritative frozen-corpus identity. The contracts repository master metadata still declares 0.2.0; codebase-memory-cli intentionally follows the published v0.2.1 release consumed by both applications rather than rewriting upstream metadata.

CP84 — COMPLETE (static only; no compile/test run): tightened integration-corpus initialization after the CP83 version reconciliation. SpecGen 0.2.1 and Agent-Workflow 0.10.0 now declare master as their pre-freeze initialization ref; agent-workflow-spec-contracts 0.2.1 declares release tag v0.2.1 because repository master still reports project version 0.2.0. Qualification instructions now require operators to verify the declared version evidence before recording the immutable frozen commit SHA.

CP85 — COMPLETE (static only; no compile/test run): made the CP84 corpus initialization contract machine-enforceable. Added scripts/qualification/validate-corpus-checkouts.py, a stdlib-only validator for repository origin, cleanliness, initialization ref, version evidence, and frozen commit identity. The corpus manifest now carries structured version_check metadata: pyproject.toml project versions for SpecGen 0.2.1 and Agent-Workflow 0.10.0, and exact tag v0.2.1 for the shared contracts repository. Qualification instructions require validation once with --allow-unpinned during initialization and again without it after immutable commit SHAs are recorded. No compilation or test execution was performed for CP85.

CP86 — STATIC REPAIR / WINDOWS BASELINE PREP: rebased from the authoritative post-CP85 source and reconciled the current benchmark integration set to SpecGen 0.2.1, Agent-Workflow 0.10.0, and shared contracts 0.2.1. Added native PowerShell benchmark and corpus-orchestration scripts plus docs/WINDOWS_BENCHMARK_RUNBOOK.md. Began static reconciliation of the 64 CLI-suite failures reported after CP85: updated CLI skill namespaces, CLI instruction markers, current hook filenames, MCP-preservation expectations, and explicit hook-surface test calls where the tests were still exercising pre-CP78 combined-install semantics. No compile or test run was performed; the next external suite run is the validation source for remaining failures.

CP87 — STATIC CLI TEST CONTRACT CLOSEOUT: consumed the external CP86 full-suite evidence (7,391 passed / 22 failed / 3 skipped; all 137 non-CLI suites green) and statically reconciled every reported CLI failure location with the CP78 split lifecycle. Hook-oriented integration tests now use the explicit hook plan/install surface; current CLI hook/script identities use the codebase-memory-cli namespace; MCP-era Gemini/Claude hook commands/scripts are asserted foreign and preserved rather than migrated; OpenClaw compaction is treated as an explicit CLI hook surface while MCP compaction remains foreign; and activation ordering tests now use an asset-owned OpenClaw instruction path rather than a hook/config path no longer touched by ordinary install. No compilation or test execution was performed for CP87; the next external scripts/test.sh run is authoritative validation.

CP86 — Windows benchmark automation and CLI contract repair

COMPLETE (static implementation): added native Windows benchmark execution (scripts/benchmark-agent-workflows.ps1) and five-repository corpus orchestration (scripts/qualification/run-windows-benchmark-corpus.ps1) while retaining exact candidate-byte provenance.
COMPLETE: reconciled the benchmark integration set to SpecGen 0.2.1, Agent-Workflow 0.10.0, shared contracts v0.2.1, and the selected managed terminal scale corpus.
EVIDENCE: external full-suite execution reduced CLI failures from 64 to 22; every non-CLI suite remained green.

CP87 — CLI test-contract closeout

COMPLETE (static repair): reconciled the 22 residual CLI assertions with the CP78 explicit-hook/MCP-preservation contract without restoring MCP ownership or implicit hook installation.
EVIDENCE: external targeted CLI execution completed 236 passes with only two Claude hook ownership/mutation assertions remaining.

CP88 — Windows RC qualification closeout

COMPLETE (static implementation; no local compile/test run): adds scripts/qualification/run-windows-rc-qualification.ps1, a top-level native-Windows RC qualification orchestrator covering exact artifact acquisition/hash verification, machine-state capture, portable smoke/canonical workflow, retained Windows guards against supplied candidate bytes, explicit recovery cycles, isolated installed-product/CLI-hook/MCP-coexistence lifecycle checks, five-repository benchmark execution, promotion-compatible evidence generation, evidence self-verification, and optional draft-release upload.
COMPLETE: scripts/setup-windows.ps1 accepts -Binary, -InstallDir, and -NoPathPrompt so qualification can install byte-identical RC candidates without rebuilding or substituting the public latest release.
COMPLETE: Windows benchmark workloads are repository-specific and frozen in BENCHMARK_CORPUS.json; the validator rejects missing workload files/fields before measurement begins.
FIXED (benchmark correctness): the PowerShell benchmark harness now fails a repository run on any required warm-up/measured operation failure and emits explicit RESULT/FAILURES.txt evidence instead of silently excluding failures from performance summaries.
FIXED (evidence isolation): corpus initialization writes the frozen manifest into the qualification evidence tree rather than dirtying the exact RC source checkout.
RECONCILED: the last two known Claude-hook tests now enforce exact current CLI command ownership while preserving MCP-era hook identities as foreign, matching the CP78 side-by-side contract.
PENDING EXTERNAL EVIDENCE: run the external CLI/full suite after CP88 and then execute the Windows RC orchestrator on the qualification host against the immutable draft artifact. No compile or test execution was performed locally for CP88.

CP89 — Windows RC qualification provenance closeout

Static/source audit only; no local compilation or test execution was performed.

external Windows qualification evidence now binds the exact release tag to its immutable source commit in addition to the Windows archive/executable SHA-256 values;
run-windows-rc-qualification.ps1 requires PowerShell 7+ for deterministic BOM-free UTF-8 JSON evidence and refuses cross-checkout SourceRoot execution;
run-windows-benchmark-corpus.ps1 resolves moving initialization branches from origin/<branch> and freezes the resulting commit before benchmarking;
the promotion workflow checks out the exact qualified tag both when verifying evidence and when publishing npm/PyPI wrappers, preventing registry publication from a later default-branch revision;
the Windows benchmark runbook now accurately describes the external frozen-manifest evidence flow and the deliberate post-qualification commit that establishes the durable corpus generation for later releases.

CP90 — Linux benchmark corpus parity

Linux benchmark-agent-workflows.sh now uses the same fail-closed required-operation semantics as the Windows benchmark harness and emits per-repository RESULT=PASS evidence only when every configured operation succeeds.
Added scripts/qualification/run-linux-benchmark-corpus.sh to clone/fetch/freeze/validate and run all five repositories from BENCHMARK_CORPUS.json in one command.
Linux and Windows corpus runners now consume the same manifest-defined operation/workload fields; repository-specific file/query/symbol/base-branch settings are no longer Windows-only behavior.
Added docs/LINUX_BENCHMARK_RUNBOOK.md. Linux is a methodology/feature-parity preflight and does not replace the native-Windows RC qualification or Windows BASELINE_ZERO evidence.

FilesExpand file tree

CLI_MIGRATION_STATUS.md

Latest commit

History