Skip to content

PROGRESS.md

Latest commit

 

History

History
826 lines (764 loc) · 49.2 KB

PROGRESS.md

File metadata and controls

826 lines (764 loc) · 49.2 KB
Raw

Progress

2026-08-17 Sparse evidence graph: CAM_Codx route

  • Added the hidden canonical knowledge-graph-query CAM_CAM path to the capability registry as a managed knowledge route with read-only/no- approval policy and graph-query report artifact.
  • Re-pinned the generated CAM_CAM manifest fixture to feature commit 6801ab008b6b239588adacbeb4eb77c988183533 (142 paths; digest 8bb56420ae57d29074ec83431f4fdf2f13f644fcaecce52e9767c99431e999fb0).
  • Added a fixed-packet test proving list-form argv and no approval, plus outcome-facing route documentation. The live validator reports 142 paths with the registry and feature checkout aligned.
  • Generated host packs are current. The initial clone-only run exposed missing sibling-checkout paths; the release gate was rerun with a bounded temporary sibling link to the verified CAM_CAM checkout.

2026-08-17 CAM_Codx full-suite release gate

  • After the CAM_CAM fixture-collection repair, the pinned manifest source was advanced to 923ea57570b45a5f8d605dddf23f10455eca3cde; manifest content and digest remain 142 paths / 8bb56420ae57d29074ec8341f4fdf2f13f644fcaecce52e9767c99431e999fb0.
  • Exact validator: Capability registry valid: 142 paths; hidden_compatibility=11, managed=129, troubleshooting_only=2.
  • Full CAM_Codx suite with the bounded sibling checkout passed 208 passed. The earlier 203 passed, 3 skipped, 2 failed clone-only result was caused solely by missing sibling CAM_CAM fixture paths and is superseded by this complete gate.

2026-08-15 control-plane crash recovery and Task 8 registry reconciliation

  • Recreated the lost temporary worktrees as isolated writable clones under /Users/o2satz/Downloads/crash814/worktrees; the SSD checkouts remain untouched.
  • Restored UPDATED_AUTONOMOUS_GOAL.md from the persisted session transcript. It is the active continuation contract for Tasks 7-14 and preserves the explicit stop at the Task 14 product-boundary decision.
  • Re-pinned the capability fixture to CAM_CAM commit dc37550ef18fe91a0e8f72acc22b2c278dafc444: 140 command paths, 16 hidden paths, and manifest digest f71dd15854dae850b36002e85a7a2ddca6bd5f59413f385b795352b87c7856c7.
  • Classified hidden canonical managed-run as a CAM_Codx-selectable record route with local_record_write risk and bounded-phase approval. It is not a competing user-facing alias, provider route, or live CAM mutation.
  • Registry tests passed 25/25; the live validator confirmed both the digest and pinned CAM_CAM revision; generated agent packs are current. The managed sandbox cannot create optional pytest cache files under Downloads, but this did not affect test results.

2026-08-15 Task 8 assess/plan managed-run composition

  • Added a fixed list-form managed-run start packet for explicit assess or plan requests with a named run ID. Packet construction hashes the target identity into the reviewed plan, pins the configured CAM runtime paths, and performs no CAM invocation, mining, provider call, or target/corpus/profile/ config mutation.
  • The packet submission boundary accepts only one fixed tuple argv and requires a zero-exit JSON-object receipt; it is dependency-injected for fixture proof and does not add a CAM runtime implementation to CAM_Codx.
  • compose_assessment now invokes the existing primary-only Development Brief builder and target inspector before it creates the bounded managed-run start packet. Direct precedent, transferable analogy, and new-hypothesis labels remain owned by the Development Brief output rather than collapsed by the router.
  • Focused verification: 51 passed across test_cam_control_plane.py and test_development_brief.py; git diff --check passed. Optional pytest-cache writes remain sandbox-blocked under Downloads only.

2026-08-15 Task 9 target-mutation plan authorization

  • Manager packets for registry policy target_code_mutation now require an existing target repository, reviewed managed-plan ID, and lowercase 64-character plan SHA-256. The target and plan identity are inside the content-addressed packet scope, so an issued single-use approval cannot be reused for another plan or target.
  • Test-first evidence: the new packet test initially failed because create accepted no plan identity; it passes after the minimal validation and scope binding change. Existing validate policy coverage now supplies an explicit target and plan identity.
  • Focused verification: 37 passed in tests/test_cam_manager.py; 88 passed across the manager, control-plane, and Development Brief tests; git diff --check passed. The only warning is the recovery sandbox declining optional pytest-cache writes under Downloads.

2026-08-15 Task 10 administrative packet routing (first slice)

  • Added prepare_admin_packet to compose the read-only control-plane route resolver with the existing manager packet constructor for knowledge, models, self-enhancement, evolution, doctor, and setup. It accepts only one of those registry intents, uses the explicit canonical command path, preserves the named run ID as workflow identity, and never invokes CAM.
  • The packet seam deliberately rejects mining. Mining remains bounded by the existing pull/mine coordinator's explicit source, corpus, model, time, cost, receipt, and delta contract until its canonical manager integration is implemented as the next Task 10 slice.
  • Test-first evidence: the new six-family packet test initially failed because the seam did not exist. It now passes with registry-declared kb search and doctor capabilities routes; the failed shorthand names proved the registry remains the selection authority.
  • Focused verification: 6 passed for the new packet test; 79 passed across router and manager tests. Optional pytest-cache writes remain sandbox-blocked under Downloads only.

2026-08-15 Task 10 mining packet preparation (second slice)

  • Added prepare_mining_packet, which validates that the pull/mine coordinator configuration matches the pinned control-plane command, corpus, config, and profile identities before it prepares a single manager mine-workspace packet. It reuses the coordinator's live argv builder, so source root, exact model, repository cap, time cap, cost cap, and budget receipt path are all approval-bound without duplicating mining logic.
  • This seam is preparation only: it creates no budget receipt, invokes no provider or CAM command, changes no corpus, and does not select or dispatch a build candidate. Execution/receipt linkage remains the next Task 10 batch.
  • Test-first evidence: the mining packet regression initially failed because the seam did not exist. It now passes and proves fixed manager argv, cost and time/repository bounds, future receipt path, no CAM invocation, and unchanged fixture corpus identity.
  • Focused verification: 1 passed for the new test; 122 passed across router, pull/mine, and manager tests. Optional pytest-cache writes remain sandbox-blocked under Downloads only.

2026-08-15 Task 10 mining receipt-to-run packet (third slice)

  • Added a fixed managed-run link-mining-receipt packet builder. It reads one existing receipt buffer, hashes that exact buffer, requires source repository identities and the explicit managed run ID, and targets only the existing CAM_CAM persistence seam with the pinned config.
  • This bridge intentionally does not submit the local-record operation, mine, create a budget receipt, mutate the corpus, or select a build candidate. The receiver validates the same absolute receipt path and SHA-256 before writing its managed-run event.
  • Test-first evidence: the receipt-link regression initially failed because the builder did not exist. It now passes and proves the fixed list-form argv, receipt digest/path, source identity, run identity, and pinned config.
  • Focused verification: 1 passed for the new link packet test; 123 passed across router, pull/mine, and manager tests; git diff --check passed. Optional pytest-cache writes remain sandbox-blocked under Downloads only.

2026-08-15 Task 11 CAM_Codx-first documentation contract (first slice)

  • Added documentation-contract tests for README, cheatsheet, program-manager guide, quickstart, and status. Normal-user documentation now starts with Use CAM_Codx to ..., names canonical cam-codx as the normal skill, and removes legacy specialized-skill invocations from those surfaces.
  • Direct CAM_CAM documentation remains positioned for troubleshooting, runtime development, recovery, regression isolation, or expert use; the contract rejects any claim that cam chat is a complete general router.
  • Test-first evidence: the contract initially failed on the old README/four- skill framing. It passes after outcome-first replacement and the legacy-skill inventory scan is empty for the five normal-user documents.

2026-08-16 Task 12 focused release gates (in progress)

  • CAM_Codx focused gate passed: 190 passed across capability registry, control-plane, manager, canonical skill, setup wizard, Development Brief, and pull/mine tests. Generated packs are current, skill frontmatter is valid, and git diff --check passed. The recovery sandbox only blocked optional pytest-cache writes.
  • The pull/mine documentation regression was repaired: its test now requires canonical Use CAM_Codx to ... language on overview surfaces and reserves detailed --source-root, --dry-run, --skip-swap, and claw.db assertions for the dedicated operator guide.
  • CAM_CAM focused gate used the actual tests/test_application_packet.py path. The planned tests/test_self_enhance.py does not exist; available self-enhancement reconstruction coverage is tests/test_reconstruct.py. The equivalent available gate produced 238 passed, 1 failed: clean CAM_CAM commit dc37550 has intentional moonshotai/kimi-k3 entries in claw.toml, but TestApprovedModelConfig excludes that ID from APPROVED_MODEL_IDS. This is an unrelated baseline policy conflict, not a passed release gate and not changed here; no model/config file was modified.
  • CAM_CAM full suite was additionally isolated with pytest -x: 469 passed, 6 skipped, 1 failed before stopping in tests/test_cag_convert.py::TestReadLanceDB::test_read_lancedb_table. claw.memory.rag_adapter.read_lancedb calls the installed LanceDB table to_pandas() path, whose underlying LanceDataset lacks that method. This is a dependency/API compatibility baseline outside the control-plane diff; it was not repaired or relabeled as a release pass.

2026-08-16 Task 12 baseline remediation

  • Corrected the stale evolution allowlist from retired moonshotai/kimi-k2.7-code to the intentionally configured moonshotai/kimi-k3; the active claw.toml profile already uses K3 for the codex agent and fallback chain.
  • Updated LanceDB import to enumerate tables with list_tables().tables and convert via table.to_arrow().to_pandas(), avoiding the installed LanceDB 0.33 to_pandas() path that delegates to a removed LanceDataset method.
  • Red evidence was the two exact Task 12 failures. Green verification: PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=src python -m pytest -q tests/test_serial_evolution.py tests/test_cag_convert.py returned 103 passed; the original CAM_CAM focused release gate then returned 239 passed. The only remaining warning was sandbox-blocked optional pytest cache creation; no runtime database, live profile, provider, or target was changed.
  • A fresh CAM_CAM full-suite probe reached 876 passed, 11 skipped before the next failure: tests/test_create_benchmark_spec.py expects monkeypatching claw.cli.ROOT_DIR to redirect _write_create_spec, but the implementation retains its root path from the monolith module and attempts to create CAM_CAM/data/create_specs. The recovery sandbox correctly denies that write. This is a distinct test-isolation/environment issue, not a regression from the Kimi or LanceDB repairs; it remains the Task 12 full-suite blocker.

2026-06-21

  • Read active GOAL.md and implementation plan.
  • Verified /Volumes/WS4TB/repo622sn/CAM_Codx is the Git-backed hub checkout on main...origin/main.
  • Verified remotes:
    • CAM_Codx: https://github.com/deesatzed/CAM_Codx.git
    • CAM_CAM: https://github.com/deesatzed/CAM_CAM.git
    • moriahcareframe: https://github.com/deesatzed/moriahcareframe.git
  • Confirmed CAM_CAM has untracked CAM_Codx_last5291pm.txt; left untouched.
  • Created non-destructive local overlay directories:
    • /Volumes/WS4TB/CAM_ALL
    • /Volumes/WS4TB/CAM_ARCHIVE/2026-06-21-pre-cleanup
  • Created clean GitHub clones under /Volumes/WS4TB/CAM_ALL/repos.
  • Documented repo inventory, local folder audit, retirement manifest, config alignment, GitHub-safe config guide, and config drift checks.
  • Rewrote README and added first-pass hub docs for architecture, Codex quickstart, Repo Necromancer workflow, MoriahCareFrame case study, status, FAQ, repo map, and launch checklist.
  • Added Codex goal templates plus Claude Code and Grok Build adapter docs and templates.
  • Added CAM_CAM backlink docs:
    • /Volumes/WS4TB/WS4TBr/CAM_Codx/CAM_CAM/README.md
    • /Volumes/WS4TB/WS4TBr/CAM_Codx/CAM_CAM/docs/showpieces/repo_necromancer/USER_GUIDE.md
    • /Volumes/WS4TB/WS4TBr/CAM_Codx/CAM_CAM/docs/integrations/CAM_CODEX.md
  • No old folders were deleted, moved, renamed, or archived.

Verification After Commits

  • CAM_Codx commit: 287ed4a docs: reorganize CAM_Codx as CAM workflow hub.
  • CAM_Codx verification report commit: 7eec8bc docs: record CAM repo reorg verification.
  • CAM_CAM commit: c911044 docs: link CAM_CAM runtime to CAM_Codx hub.
  • Pushed CAM_Codx main to GitHub through 7eec8bc.
  • Pushed CAM_CAM main to GitHub through c911044.
  • MoriahCareFrame had no changes to push and remained at a82e42c.
  • CAM_Codx git diff --check: passed.
  • CAM_CAM python -m pytest -q tests/test_repo_necromancer.py: 6 passed.
  • MoriahCareFrame PYTHONPATH=src python -m pytest -q: 5 passed.
  • MoriahCareFrame sh scripts/smoke.sh: passed and left git status clean.
  • CAM_ALL verify-all.sh: passed for clean clones.

Final Public Cleanup Continuation

  • Re-read active GOAL.md; scope is now the final public repo cleanup and fresh-clone proof pass, not just the initial hub reorganization.
  • Verified current pushed heads before cleanup:
    • CAM_Codx: 23ad6e555127dc2856eec5070c91f2c09c04b238
    • CAM_CAM: c9110447abff2047a8c4df7021679a0847fb151e
    • moriahcareframe: a82e42cedd2f70479d44f92bd2dcab7277f86168
  • Confirmed current dirty state before edits:
    • CAM_Codx: modified GOAL.md.
    • CAM_CAM: untracked CAM_Codx_last5291pm.txt, left untouched.
    • MoriahCareFrame: clean.
  • Identified stale current-state references in status/report/inventory docs and public config docs that mislabeled tracked public-safe CAM_CAM/claw*.toml defaults as local-only.
  • Started cleanup classification with generated CAM_CAM batch outputs, stale launch reports, and dated coverage baseline as low-risk public Git removal candidates.
  • Committed and pushed CAM_CAM cleanup: 9a9d71ade8f6766c8fb564051b2baa308d9abfd1 chore: remove stale public cleanup artifacts.
  • Committed and pushed CAM_Codx cleanup manifest batch: 7a142e3e4957f270c5179693330030dabb9cbfd0 docs: record final public cleanup manifest.
  • Created fresh-clone proof directory: /Volumes/WS4TB/CAM_ALL/clone_proofs/2026-06-21-public-cleanup-104001.
  • Fresh-clone verification passed for:
    • CAM_Codx: git diff --check, JSON validation, TOML validation, required docs/templates, and stale planned-status scan.
    • CAM_CAM: python -m pytest -q tests/test_repo_necromancer.py, git diff --check, and absence checks for removed stale artifacts.
    • moriahcareframe: PYTHONPATH=src python -m pytest -q, sh scripts/smoke.sh, and git diff --check.
  • Updated /Volumes/WS4TB/CAM_ALL/repos/* to current origin/main and verified /Volumes/WS4TB/CAM_ALL/scripts/verify-all.sh passes.
  • Narrow high-entropy secret scan found only CAM_CAM test fixture strings; no tracked .env, database, sqlite, pem, or key files were present in fresh clones.
  • Post-claim validation found stale final-head/status wording in CAM_Codx docs. Updated final status/report/manifest wording so the current pushed CAM_Codx proof head is ef1e20f14870bfcc55d0a508e06b1726e6f02e8f, later containing commits are verified by git rev-parse HEAD, and tracked public-safe CAM_CAM/claw*.toml defaults are not mislabeled as uncommitted local-only files.

2026-06-23 CAM Agent Packs Goal

  • Replaced the prior final-public-cleanup GOAL.md with a new autonomous completion contract for CAM Agent Packs.
  • Locked the ownership decision: CAM_Codx remains the main workflow hub; CAM_CAM remains the runtime/MCP core; Claude Code, Gemini, and Grok Build are generated host-specific packs, not separate product forks.
  • Included proof gates for a shared capability contract, deterministic pack generator, host-specific packs, tests, CAM_CAM runtime verification, official host-doc rechecks, and no-secret hygiene.
  • Current local verification for this goal-authoring step: git diff --check passed after editing GOAL.md.
  • Implemented agent-packs/contract/cam_agent_capabilities.json with the CAM runtime ownership model, checked external docs, required host packs, and 19 CAM CLI/MCP capabilities.
  • Added tools/generate_agent_packs.py, generated agent-packs/contract/CAPABILITY_CONTRACT.md, docs/AGENT_PACKS.md, and generated packs for Claude Code, Gemini, and Grok Build.
  • Added tests/test_agent_packs.py covering required capabilities, host-pack required files, generated-output freshness, JSON example parsing, and no-secret/no-local-DB hygiene.
  • Updated README, status, repo map, and integration docs for Claude Code, Gemini, and Grok Build.
  • Rechecked CAM_CAM runtime ownership from src/claw/mcp_server.py, src/claw/tools/schemas.py, docs/MCP_INTEGRATION_GUIDE.md, and cam premine CLI wiring.
  • CAM_Codx verification passed: python -m json.tool agent-packs/contract/cam_agent_capabilities.json, python tools/generate_agent_packs.py --check, python -m pytest -q tests/test_agent_packs.py with 6 tests, and git diff --check.
  • CAM_CAM runtime verification passed: python -m pytest -q tests/test_tool_schemas.py tests/test_integration_wiring.py with 74 tests, and git diff --check.
  • Host CLI availability checks: claude mcp list exits 0 but currently lists only Claude Google connectors that need authentication; CAM is not installed there yet. gemini mcp list exits 0 and reports no MCP servers configured. grok inspect exits 0 and reports no project MCP servers configured for this CAM_Codx checkout.
  • Standardized the generated Claude Code, Gemini, and Grok Build pack READMEs around the same setup and test sections: Quick Start, Configure CAM MCP, Verify Discovery, Smoke Test, CAM Capabilities, Safety Policy, and Files.
  • Added one executable smoke.sh per host pack and removed the Grok-only headless-smoke.sh name so every pack has the same test entrypoint.
  • Added the uniform setup/test flow and smoke-script links to the CAM_Codx landing README and aligned the Claude, Gemini, and Grok integration docs.

2026-07-06 CAM_Codx Setup Wrapper

  • Designed the third setup option for Codex/CAM permissions: generate a narrow cam-codx wrapper instead of asking users to grant broad writes to the CAM_CAM install directory.
  • Added design and implementation-plan docs:
    • docs/plans/2026-07-06-cam-codx-setup-wrapper-design.md
    • docs/plans/2026-07-06-cam-codx-setup-wrapper.md
  • Extended tools/cam_setup_wizard.py with wrapper generation and explicit --wrapper-cam-cam, --wrapper-db, --wrapper-config, and --wrapper-env arguments.
  • Updated setup documentation to explain the Codex approval prefix: <CAM_HOME>/scripts/cam-codx.
  • Current targeted verification: python -m pytest -q tests/test_cam_setup_wizard.py passed with 8 tests.
  • Full targeted verification passed: python -m pytest -q tests/test_cam_setup_wizard.py tests/test_agent_packs.py passed with 17 tests, python tools/generate_agent_packs.py --check passed, and git diff --check passed.
  • Live wrapper proof for this machine: /Volumes/WS4TB/codxswarm/scripts/cam-codx status passed and reported all four agents executable; /Volumes/WS4TB/codxswarm/scripts/cam-codx stats passed and reported 2,474 active methodologies.
  • Added the final novice setup step: tools/cam_setup_wizard.py --install-codex-skill now installs templates/skills/cam-codx-setup into the user's Codex skills directory and auto-detects side-by-side clones at <CAM_HOME>/CAM_CAM.
  • Verification for the novice setup path: python -m pytest -q tests/test_cam_setup_wizard.py tests/test_agent_packs.py passed with 19 tests, and python tools/cam_setup_wizard.py --cam-home /Volumes/WS4TB/codxswarm --skip-clone --install-codex-skill --codex-home /tmp/cam_codx_fake_codex_home --non-interactive created the wrapper and installed the skill into the temporary Codex home.
  • Added the novice startup explanation to docs/QUICKSTART_CODEX.md, covering what Codex, CAM_Codx, CAM_CAM, claw.db, and cam-codx do; the exact clone/install/setup/start-Codex flow; and common CAM_Codx use cases.

2026-08-09 Post-Mining Cleanup And Augmentation Audit

  • Preserved and tracked the validated GOAL_CAM_SUBSCIBED.md contract.
  • Corrected active CAM_Codx runtime references from the older WS4TBr checkout to /Volumes/WS4TB/repo622sn/CAM_CAM in the goal, capability contract, architecture, repo map, and config guidance.
  • Classified the successful 80-finding mining batch without additional model calls or provider spend.
  • Recorded the audit in docs/reports/2026-08-09-post-mining-cam-augmentation-audit.md.
  • CAM_CAM cleanup created a tracked database registry, retained but ignored SQLite WAL/SHM files, preserved intentional Kimi K3 configuration, and federated Go, misc, TypeScript, and Rust ganglia.
  • Verified root corpus status: 2,668 methodologies, 215 source repositories, four configured siblings, and SQLite integrity ok for all five databases.
  • Focused CAM_CAM verification passed: 161 tests, with two known non-fatal aiosqlite event-loop shutdown warnings recorded as separate audit debt.
  • Chose runtime identity preflight and reviewed adoption manifests as the next P0 implementation batch. No automatic self-enhancement or CAG rebuild was performed in this audit.

2026-08-10 CAM_Codx Program Manager

  • Approved and committed the CAM_Codx program-manager design and implementation plan in docs/plans/2026-08-10-cam-codx-program-manager-{design,}.md.
  • Added tools/cam_manager.py, a fixed-operation, no-shell packet executor with content-addressed scope, short-lived single-use approvals, and digest-only execution receipts.
  • Added templates/skills/cam-codx-swe/SKILL.md and setup-wizard installation of both the setup and routine SWE skills.
  • The routine path explicitly separates ordinary CAM recall from mining, provider spend, model promotion, and self-enhancement swap.
  • CAM_CAM tournament lineage hardening passed its focused 88-test gate.
  • A manager-approved, supervised self-enhancement run executed one task against a disposable CAM_CAM copy with --max-tasks 1 --skip-swap. The generated candidate changed src/claw/memory/auto_fix.py, but its focused regression failed 5 tests, so the candidate was rejected and no live source, database, profile, or configuration was swapped. This is the intended fail-closed outcome; the candidate is not a proposed CAM change.
  • The manager execution receipt recorded return code 0 for the CAM CLI, while CAM's own validation result is the acceptance authority. The disposable copy is retained only long enough to inspect the failed candidate and will not be published.
  • Final runtime audit corrected the model-promotion allowlist to CAM_CAM's current models set, models rollback, and models profile use commands; focused tests cover all three approval-required prefixes.

2026-08-10 SWE Development Brief

  • Added the approved Development Brief contract, target inspector, primary-only CAM recall adapter, direct/analogy/hypothesis labels, and explicit CLI.
  • Default operation reads the named target and supplied primary corpus only; it does not write a target file, record retrieval usage, execute target tests, mine, invoke a provider, or query sibling corpora.
  • Added the cam-codx-development-brief skill and setup-wizard installation alongside the setup and routine SWE skills.
  • Named local source roots are validated under an approved parent and render only a later scan-only proposal. Missing configured sibling databases render a relocation gate rather than triggering a broader search.
  • Focused Development Brief and setup-wizard tests are recorded by the final cross-repository verification task; no live corpus was queried for this work.
  • Final CAM_Codx verification in the isolated Development Brief worktree: python -m pytest -q -p no:cacheprovider tests passed with 44 passed; python tools/generate_agent_packs.py --check and git diff --check passed.
  • The companion CAM_CAM worktree verification passed 78 focused tests across test_read_only_brief_query.py, test_tool_schemas.py, and test_integration_wiring.py; python -m claw.cli brief-query --help passed. The read-only proof uses a synthetic fixture database only, not a live corpus.

2026-08-11 Development Brief Documentation Alignment

  • Updated the landing README, Codex quickstart, CAM cheat sheet, program-manager guide, and status page so users can choose between the Development Brief, routine SWE skill, and approval-gated manager workflow without confusing their boundaries.
  • The setup documentation now lists all three installed skills. The early new-project and continue/rescue prompts are copy-pasteable from the quickstart and cheat sheet.
  • Re-verified python -m pytest -q -p no:cacheprovider tests (44 passed) and python tools/generate_agent_packs.py --check; git diff --check passed.
  • This documentation pass changed no CAM runtime, database, model profile, configuration, or target repository.

2026-08-11 Pull Mine Directory Skill

  • Added tools/cam_pull_mine_dir.py and the installable cam-codx-pull-mine-dir skill for one explicit directory-level CAM cycle.
  • The coordinator defaults to /Volumes/WS4TB/waswiki/repos2mine/repo622sn but accepts --source-root for another operator's repository directory.
  • Eligible Git repositories receive only git fetch origin then git pull --ff-only; dirty, conflicted, detached, no-upstream, fetch-failed, and non-fast-forward repositories are reported without blocking later ones.
  • The mining command is list-form, pins both CAM database environment variables to one claw.db, runs scan before live mining, requires a paired exact model and hard cost cap, uses --changed-only --no-tasks, and records only bounded redacted command evidence.
  • The workflow reports read-only corpus integrity/deltas and ledger provenance. It can dispatch at most one manager-backed supervised --skip-swap candidate only after the five-findings/two-repository/repeated-gap gate. A swap, model/profile change, rollback, source edit, or live config change is outside this skill.
  • The repeated-gap conclusion is an explicit --repeated-pattern-or-gap attestation because the current corpus/ledger do not prove that semantic conclusion on their own.
  • Verification was fixture-only: python -m pytest -q -p no:cacheprovider tests passed with 86 passed in 1.09s; python tools/generate_agent_packs.py --check, the new skill validator, and git diff --check exited successfully. No live repository pull, corpus mining, provider call, database/ledger update, or candidate was executed while implementing this feature.

2026-08-12 CAM_Codx Unified Control Plane Design

  • Audited the current CAM_CAM Typer command tree, hidden aliases, nested command groups, CAM_Codx tools, installed skill templates, setup behavior, and CAM-SEQ persistence surfaces.
  • Confirmed that cam chat currently executes the mining route but explicitly reports create/build and enhance/fix routing as not wired.
  • Confirmed that setup currently installs four overlapping skills while a separate semantic session-router artifact also exists.
  • Approved CAM_Codx as the normal user-facing manager for every CAM_CAM capability. Direct CAM_CAM usage is retained for troubleshooting, runtime development, recovery, and regression isolation.
  • Approved six everyday SWE intents: assess, plan, build, fix, verify, and record.
  • Approved one canonical future cam-codx skill, one machine-readable capability registry, hidden compatibility aliases, and explicit approval classes for corpus writes, provider spend, code mutation, promotion, and CAM live swaps.
  • Approved a source-to-outcome SWE Run backed by the existing CAM-SEQ task_plans, component/application packets, pair/landing/outcome events, and run connectomes rather than a parallel database.
  • Recorded the approved design in docs/plans/2026-08-12-cam-codx-control-plane-design.md and made it the active GOAL.md contract.
  • This checkpoint is documentation and design only. The capability registry, canonical skill, router, migration, and MatrAIx/SESA proof remain unimplemented.
  • Added the exhaustive current command/skill audit at docs/CAM_CAPABILITY_AUDIT_2026-08-12.md.
  • Added the TDD implementation plan at docs/plans/2026-08-12-cam-codx-control-plane.md.
  • Confirmed that agent-packs/contract/cam_agent_capabilities.json is the existing registry foundation to extend; the plan does not create a second competing capability registry.
  • Added current-versus-approved-target notices to the landing README, cheatsheet, program-manager guide, quickstart, and active IMPLEMENT.md.
  • Documentation/plan checkpoint verification passed: 86 passed across agent packs, setup wizard, manager, Development Brief, and pull/mine coordinator; python tools/generate_agent_packs.py --check and git diff --check also passed.

2026-08-12 CAM runtime capability registry

  • Extended the existing agent capability contract to schema 2.0; it remains the single registry for host capabilities, the 13 approved CAM_Codx workflow intents, and every CAM_CAM command/group route.
  • Captured an independent schema-version-1 manifest snapshot from CAM_CAM Task 1 and validated exact coverage of 139 paths: 126 managed, 2 troubleshooting-only, and 11 hidden compatibility aliases.
  • Added strict validation for missing, unknown, duplicate, multiply classified, kind-mismatched, hidden-mismatched, and incomplete-policy routes. Hidden compatibility commands remain callable but are omitted from generated normal and direct-runtime references.
  • TDD RED produced 9 failed, 9 passed before the schema, validator, and generator support existed. Focused GREEN passed 18 registry/agent-pack tests; the live snapshot validator, generator --check, and git diff --check also passed. These checks made no provider call and did not change a CAM database, model profile, or runtime configuration.
  • Specification review found that shape-only policy validation had mislabeled several provider-backed, artifact-writing, service-starting, schema-writing, and target-executing commands as read-only. The follow-up audit corrected those route policies conservatively, added closed policy vocabularies and approval/flag invariants, and added named runtime-boundary regressions.
  • The second policy-truth pass also corrected synergies, kb brains, and self-enhance status: each opens SQLite through an engine path that may create files or set WAL state, so none is represented as side-effect-free.
  • Quality review separated the 11 genuine hidden aliases from four hidden canonical operations (evolution approve, govern, mine-report, and prism-demo) and added explicit alias targets. Hidden canonical operations remain manager-routable but are absent from normal choices.
  • Added compound approval_classes plus a risk/side-effect/default/flag compatibility matrix. Corrected community publish/import external-network policies and restored read-only truth for mine-report.
  • Pinned the independent manifest fixture to CAM_CAM commit 5075645 and its content digest. The cross-repo gate now generates the live manifest from an explicit runtime checkout, compares its digest and exact Git revision, and then validates registry coverage.
  • Final specification review aligned all alias safety tuples with their canonical targets and made parity a validator invariant. It also corrected the report/delta/synergy pairs to local initialization/write policy with no provider spend, matching their current runtime implementations.
  • The final quality pass extended alias parity to workflow routes and artifact promises, added a complete risk/default-mode matrix, and made malformed manifest values and provenance fail with controlled diagnostics.

2026-08-13 read-only CAM_Codx control-plane router

  • Added a typed, registry-backed cam_control_plane.py plan surface for the six everyday SWE intents and seven explicit administrative families.
  • Planning requires pinned absolute identities for the target, CAM executable, database, config, and optional model profile plus optional SWE Run and mining receipt references. Unknown intents, unresolved/non-executable paths, colliding runtime identities, intent/operation mismatch, and registry gaps fail closed.
  • The JSON result and human status card expose the goal, selected route, target, memory mode, operation write boundary, provider-spend possibility, mining status, approvals, and next action. Planning does not import or invoke CAM_CAM and explicitly records operation_executed=false.
  • TDD RED produced 24 expected failures while the module was absent. The first specification review found unsafe lexical route selection, no proof that claw.toml and the explicit database named the same identity, and incomplete no-write snapshots. The hardened router now uses explicit operation words with intent-specific safe fallbacks, validates the TOML database binding, and hashes structural metadata, symlinks, empty directories, .git, and SQLite sidecars. An executable tripwire proves the CAM command is never invoked during planning.
  • GREEN now passes 28 focused tests and 63 combined control-plane, capability-registry, and agent-pack tests, plus CLI help and git diff --check. No CAM operation, provider call, database mutation, target write, model-profile change, or runtime-config change occurred.
  • Quality review found that unordered free-text matching could interpret a negated mention as a mutating operation. Non-default operations now require the explicit operation field; ordinary text always uses the one default command declared by the capability contract for that intent. The registry validator proves each default is one visible managed canonical command in the matching family.
  • Full selected-route/config shape checks now turn malformed registries and TOML database tables into controlled exit-code-2 errors without tracebacks. The hardened combined suite passes 69 tests and generated agent-pack drift checks pass.

2026-08-13 contract-driven CAM approval manager

  • Replaced the duplicated manager prefix, phase, and read-only maps with a strict adapter over the shared capability contract. All 110 executable managed canonical CAM commands, including the four hidden canonical advanced operations, now receive a fixed list-form argv prefix and the contract's CAM_Codx phase. Groups, troubleshooting-only commands, and hidden compatibility aliases cannot be selected as canonical manager operations.
  • Retained the existing hyphenated manager operation names only as translations to exact canonical command paths. They carry no separate risk, side-effect, approval, spend, configuration, or promotion policy.
  • Packets now bind the canonical operation policy and trusted contract digest into their content-addressed scope. Preparation and execution fail closed on missing/malformed contracts, changed contract bytes, changed packet fields, changed policy, invalid prefixes, or an unexpected contract identity.
  • Preserved secret rejection, digest-bound approvals, expiry, single-use consumption, secure receipts, and shell=False execution. Extended budget validation to reject NaN and infinities as well as negative values.
  • TDD RED failed at collection because the new contract adapter did not exist. GREEN passes 56 focused manager/registry tests and 143 combined manager, registry, pull/mine, control-plane, and agent-pack tests. Generator drift, CLI help, and git diff --check also pass. Tests use fixture wrappers only; no live CAM operation, provider call, corpus write, target mutation, promotion, or configuration change occurred.
  • Specification review found three fail-closed gaps: the manager's partial contract validation could accept contradictory spend/approval policy, a recomputed packet could substitute its executable wrapper, and approval reuse was guarded by a check-then-append race. The manager now calls the same strict policy validator as the registry gate, requires the trusted wrapper identity again at execution, and atomically claims each approval through an exclusive mode-0600 consumption record before dispatch.
  • New regressions prove an unsafe provider_spend=true/approval=none contract is rejected, a recomputed wrapper substitution never runs, and two concurrent consumers produce exactly one success. The hardened combined suite passes 146 tests; direct-script CLI help, generator drift, and git diff --check pass.
  • Quality review then found that a wrapper could be replaced in place at the same path and that --dry-run burned the single-use approval. Packet scope now binds the wrapper SHA-256 and verifies it both during validation and immediately before dispatch. Approval validation is separate from atomic consumption, so a dry run checks the supplied approval but leaves it usable for the later real execution.
  • In-place replacement and dry-run-then-execute regressions pass. Final focused manager/registry verification is 61 passed; the combined Task 4 compatibility surface is 148 passed, with direct CLI help, generator drift, and git diff --check green.

2026-08-14 canonical CAM_Codx skill

  • Initialized the canonical cam-codx skill with the official skill-creator scaffold and UI metadata, then replaced the placeholder with one concise normal-work router and four one-level progressive references: SWE, knowledge/mining, administration, and safety/approvals.
  • The required flow reads target truth first, resolves exact runtime identities, plans before execution, displays the route/write/spend/mining card, uses the contract-driven fixed-operation manager, makes evidence selection/rejection and landing explicit, verifies, and records only the proved outcome.
  • The SWE playbook covers early new-project recall, in-progress continuation, rescue, mitigation, troubleshooting, re-development, dissimilar donor evidence, landing maps, and failed-verification handling. It reuses the existing Development Brief and pull/mine coordinator rather than copying their implementation.
  • The skill exposes all six everyday intents and seven administrative families while stating no implicit mining, no implicit promotion, separate provider/target/promotion/live approvals, and direct CAM_CAM only for troubleshooting/runtime development/recovery.
  • TDD RED produced seven failures against the generated placeholder. GREEN is 7 passed; the repo frontmatter validator reports zero failures, the official quick validator reports Skill is valid!, and git diff --check passes. No CAM command, provider call, corpus write, target mutation, promotion, or configuration change occurred.
  • Specification review found that the first draft still named two legacy skills as dependencies. The canonical playbooks now call tools/development_brief.py and tools/cam_pull_mine_dir.py directly, with pinned arguments and an explicit mining dry run. Tests forbid references to the legacy skill entrypoints, preserving one normal CAM_Codx skill.
  • Quality review found that approval looked automatic, non-default operation and argument syntax was incomplete, and the mining example omitted the bounds promised by its prose. The skill now requires authorization covering every declared class, records the actual source with --approved-by, and forbids Codex from self-authorizing spend, configuration, promotion, or live mutation. Planner and manager examples show explicit canonical operations and list-form JSON arguments.
  • The mining dry run now pins profiles/local defaults, exact provider/model, repository count, duration, and cost. A live run may remove only --dry-run from the reviewed command; any changed value requires new review.

2026-08-14 canonical skill setup migration

  • Changed default Codex skill installation from four specialized skills to the single canonical cam-codx package. Existing legacy entries are detected and reported but remain untouched unless migration is explicitly requested.
  • Added --migrate-codex-skills, valid only with canonical installation. It moves exactly cam-codx-setup, cam-codx-swe, cam-codx-development-brief, cam-codx-pull-mine-dir, and cam-codx-session; unrelated skills are never selected.
  • Each explicit migration creates a timestamped mode-0700 backup and updates a mode-0600 restore.json after every move. The metadata records original and backup paths and remains partial if a move fails, so migration never silently deletes or strands already moved entries without a recovery map.
  • Updated the setup skill to document canonical-only installation, explicit migration, backup, restoration metadata, and the unrelated-skill boundary.
  • TDD RED failed at import before the migration helpers existed. GREEN passes 22 setup/canonical-skill tests; wizard help and git diff --check pass. Verification used temporary Codex homes only and did not alter installed user skills, CAM runtime state, a corpus, model, or configuration.
  • Specification review found that legacy migration ran before canonical installation and that the partial-failure branch lacked proof. Setup now installs cam-codx successfully before moving any legacy entry. A forced second-move failure produces a controlled error pointing to restore.json, leaves the first move recoverable, leaves the unattempted legacy skill in place, and records status=partial. The selected suite is now 117 passed.
  • Quality review found two crash-safety gaps: replacing an installed canonical skill deleted it before the new copy was known-good, and the migration journal learned each move only after it happened. Canonical updates now copy to a sibling staging directory, durably journal and move the old install to a mode-0700 backup, atomically replace it, and restore it if the swap fails. The report exposes the backup and mode-0600 restore metadata.
  • Legacy migration now durably writes the complete original-to-backup plan before its first move and records a per-entry state after each one. New regressions cover copy failure, swap rollback, successful canonical backup, second-move failure, and interruption after a move but before its state update. Focused setup/skill verification is 27 passed; the selected Task 6 compatibility surface is 121 passed, with wizard help and git diff --check green. All tests used temporary Codex homes.

2026-08-16 Read-only model-comparison packet

  • Registered CAM_CAM's new canonical models benchmark compare command from runtime commit 5a9b8098b4fbb72532a4c719c85b16254bbb3db7. The pinned manifest now has 141 paths and digest daa8e3d6ae946d55f71831468fc204b0d696758e15e515a4fb8b397305f06123.
  • The benchmark-compare manager alias constructs fixed list-form argv for the existing three-stage comparison service. It is read_only, has no approval class, and cannot spend, promote, select, or edit configuration.
  • Test-first evidence: the packet test initially failed because the alias was unsupported. After registry, fixture, and generated-pack updates, 75 passed across manager, registry, and agent-pack tests; the live CAM_CAM manifest and exact pinned revision validated successfully. Optional pytest-cache writes remain sandbox-blocked in the recovery worktree.

2026-08-18 Tasks 10-13 release audit

  • Corrected the active control-plane plan to invoke the real CAM_CAM test path tests/test_application_packet.py; the stale tests/planning/... path now appears only in the historical continuation contract's explanation.
  • Current CAM_Codx focused Task 10-13 gate passed 179 tests. Generated packs, skill frontmatter, the official skill validator, and git diff --check passed. CAM_Codx's current full suite passed 208 tests.
  • Ran a temporary Codex-home proof against the current templates: the only newly installed skill was cam-codx. A fixture assess plan returned operation_executed=false, and target/database/config/model-profile identity hashes were unchanged (identity_unchanged=true). No CAM command, provider, corpus write, target mutation, or profile/config change occurred.
  • CAM_CAM's current full suite completed with 4457 passed, 22 skipped, 9 failed. The nine failures remain the documented recovery-checkout artifact-write permission failures (seven) and inaccessible real-ganglia databases (two); the CAM_CAM graph/adapter gate remains green at 22 passed.
  • Cross-repository registry validation reports 142 paths with 11 hidden compatibility, 129 managed, and 2 troubleshooting-only routes. Both feature branches and both main refs are pushed and clean. Next action is the required requirement-by-requirement completion audit; Task 14 remains outside this goal.

2026-08-18 Documentation synchronization and Task 14 contract

  • Synchronized active CAM_Codx status, goal, implementation, manager, audit, publication, and continuation surfaces with the Tasks 1-13 release audit. Historical 2026-08-12 capability statements and crash-recovery paths now carry explicit historical/superseded notices instead of being rewritten.
  • Added GOAL_TASK_14_MATRAIX_SESA.md as a discovery-first successor contract. It does not invent a MatrAIx/SESA target path or authorize live import, provider spend, target mutation, model promotion, or deployment. One meaningful product-boundary checkpoint is required before those phases.
  • Documentation gate: CAM_Codx documentation/skill/pack tests 22 passed, generated packs --check passed, CAM_CAM graph/import gate 22 passed, Task 14 required-section/whitespace validation passed, and both git diff --check checks passed.
  • Historical assumption at the documentation-sync checkpoint: the exact MatrAIx/SESA checkout, immutable revisions, ownership, license, privacy boundary, provider policy, and landing scope were not yet resolved. The subsequent source preflight resolves the paths/revisions and retains the remaining rights and product-boundary limitations below.

2026-08-18 Task 14 source preflight and canonical-root consolidation

  • Fast-forwarded the original /Volumes/WS4TB/waswiki/CAM_Codx and /Volumes/WS4TB/waswiki/CAM_CAM main checkouts to the published recovery heads, then pushed both main and feature refs. The Downloads worktrees are now recovery references only.
  • Preserved the one existing CAM state set at /Volumes/WS4TB/waswiki/CAM_CAM/{claw.db,claw.toml,.env}. The database, config, and secrets file were not copied, swapped, or modified.
  • Resolved source identities: MatrAIx-Persona-8B at 3202a0bf6134776735c4ab4d50de79be8c6a5e8b with an MIT root license; SESA at 74de5d77a19774cfba53d6950d47633a2d632430 with no root license and mixed nested third-party notices. The evidence is in docs/reports/2026-08-18-task14-source-preflight.md.
  • No project code, provider, model, dataset, CAM database, .env, or live configuration was executed or changed. SESA adaptation/redistribution remains blocked until its rights boundary is established; read-only assessment remains allowed.

2026-08-18 Task 14 CAM/CAM_Codx proof scenario

  • Created docs/plans/2026-08-18-task14-cam-proof-scenario.md and ran its read-only phases against the pinned MatrAIx and SESA sources.
  • Both session preflights were clean with one expected warning: the canonical CAM directory cannot create WAL sidecars. Both Development Briefs completed without mutation; MatrAIx had no repository truth/evidence, while SESA had no truth files and visible NotImplemented markers.
  • Both CAM_Codx plans resolved to assess -> brief-query, read-only, no writes, no mining/provider spend, no approval, and operation_executed=false. Explicit CAM primary-only brief queries returned results: [].
  • A malformed DB/config identity was rejected fail-closed; the corrected SESA plan passed. Runtime config/database hashes remained unchanged, and no target, database, config, provider, model, dataset, or source code changed.
  • Reported verdict: safety/orchestration proof passes; useful mine-to-build and product proof remains unproven. Next action is the independent candidate ledger and clean-room landing map.