Repository documentation

1.1 SDLC audit delivery

• Status: active
• Owning queue item: SDLC-AUDIT-001
• Completion / archival evidence: pending while SDLC-AUDIT-001 remains open

The user scheduled this program for 1.1 on 2026-09-10. Goals 1–6 in PROJECT.md own onboarding, provider-neutral generation and independently verifiable release evidence. The following order is the execution contract; checked source code alone does not close an evidence gate.

Expanded integration scope

RELEASE-INTEGRATION-001 records the maintainer's subsequent directive to complete all outstanding roadmap work, map every unresolved TODO into 1.1, integrate current and outstanding branches to main, and address/annotate/close all open PRs and issues. The first reconciliation found 94 open or partial active-queue items and four closed parents with unchecked acceptance. All 98 unresolved parents are now explicitly targeted at 1.1.0; the four inconsistent parents were reopened. Existing completed items and historical target text remain available as evidence, not fresh promises to recreate old releases. The prior 21-item projection below is a historical checkpoint, not the current scope limit. Literal TODO/FIXME occurrences in detector patterns, negative test fixtures and explanatory prose are not unimplemented product tasks.

Current remote main includes library PR #370. PR #372 addresses reviewed library-kind promotion and is still running hosted tests; PR #364 records the HTML Phase 0 landing. Integration must preserve these contributions and qualify the combined revision.

Assessment at program start

These are engineering judgments on the inspected working tree, not benchmark scores or certification. A = strong demonstrated capability; B = sound with material gaps; C = usable but inconsistent or incompletely verified; D = fragile or largely unproven. Plus/minus indicates position within a band. Potential is graded separately from demonstrated results.

The strongest concept is an evidence-preserving lifecycle for specification-led software. “Markdown plus an agent” alone is not a defensible research contribution. The testable contribution is whether specification authority, independent acceptance and retained evidence reduce regression, repair effort and migration cost across repeated changes. That is promising and not yet demonstrated comparatively.

Architectural defense: strict records and many small contracts impose real complexity, but they make invalidation and provenance inspectable. Preserve them. Consolidate duplicated packaging/pin mechanics and long adapter responsibilities incrementally; do not rewrite the architecture to improve a style score.

Research evaluation required for 1.1 claims

Use a declared, held-out task set covering new construction, brownfield changes, cross-language regeneration and dependency/API changes. Compare the same model/tool budget in plain coding-CLI, specification-only, full Literate AI and targeted ablations (without graph retrieval, without retained context, without independent oracle). The ablations are isolated experiments, never weakened production gates. Keep private acceptance material inaccessible to generation.

Record exact model/provider/version, task/spec/source/oracle identity, environment, repetition, success/failure, reason, tokens, wall time, estimated cost with pricing provenance, human interventions and regressions. Analyze matched task sets; report failures and missing trials. Report uncertainty and per-task paired differences instead of counting repeated trials as independent tasks. Do not call unrun trials a pass.

Primary-source context: Anthropic's Building effective agents supports measuring simple baselines before increasing orchestration. OpenAI's analysis of SWE-bench Verified limitations identifies evaluation validity and contamination concerns; this is why the proposal uses fresh held-out tasks and independently reviewed acceptance. Neither source evaluates Literate AI.

Audit correction after source recheck

The original report misread docs/roadmap/active-work.md:7064#.... documentation.py:553 emits the source line of a broken Markdown link. No line-number pin exists here. Renaming PLUGIN-001 left its old heading slug in another link. Repair that link and preserve normal heading-link validation; introducing a new locator protocol is unjustified. Also, the previous “9 passed with 1 skip” phrasing means 9 tests ran: 8 passed and 1 skipped.

Ordered release work

Design review checkpoint

ADR 0039 is Accepted by explicit maintainer approval. Its library, adoption, plugin and execution-support boundaries now authorize dependency-ordered implementation. Local plugin packaging remains a prototype pending complete routed-skill packaging and provider qualification. Acceptance does not authorize model spending, publication or weaker release gates.

Reproducible research evaluation

python3 scripts/summarize_research_trials.py TRIALS.json --baseline plain-cli validates a literate-ai/research-trials@1 document with a trials array. Each row names task, trial, variant, model, environment, specification_identity, oracle_identity, boolean passed, and nonnegative integer elapsed_ms, tokens, cost_microusd, human_interventions. Costs must come from the retained run's pricing provenance; this helper does not look up pricing or invent missing costs. Use the same task/trial pairs and model/environment/authority in each comparison. Keep raw evidence alongside the input and never include secrets or hidden oracles.

The summary reports actual totals and paired task-weighted differences, with a deterministic task-level bootstrap interval. Repeated trials of one task do not inflate the independent sample size. Fewer than 20 tasks are flagged as a small sample; a single task has no interval. These summaries describe supplied data, not authenticated acceptance or demonstrated product superiority. The unit-test fixture is synthetic and supplies no evidence of efficacy. Live experiments remain unmeasured until their retained records exist.

Current tracker inventory

Read-only survey on 2026-09-10 found nine open issues and two reviews. Ownership: #361 CANONICAL-JSON-001; #367 ADOPT-PROMOTION-001 (existing PR #369); #368 ADOPT-REFRESH-001; #365 ADOPT-MONOREPO-001; #295/#296 OBSERVE-HTML-001; #280 LIBRARY-ARTIFACT-001; #279 SUBMODULE-ORCH-001; #251 RELEASE-017/GOLDEN-PATH-001. PR #364 is the existing observability planning review. Do not duplicate those implementations or infer that an open milestone proves release readiness. Tracker milestone/disposition changes and final closure remain explicit release operations.

Correction on #251: GOLDEN-PATH-001 is already completed for 0.10; the open issue is the separate historical upgrade obligation in RELEASE-017. A 1.0.1 → 1.1 smoke must not falsely close a promised 0.8.4 → 0.9 migration proof. Resolve that historical disposition explicitly. v1.0.1 was published on 2026-09-10 and is the current predecessor.

PYTHONPATH=src:. python3 scripts/summarize_release_roadmap.py 1.1 gives a deterministic, non-writing projection of explicitly targeted queue items and their unchecked acceptance. It rejects duplicate IDs and false closed checklists; an empty selection is never ready. It is a local queue view, not a replacement for the live contribution sweep, historical obligations, or release-policy gates.

Intelligence evidence and experiments

tests.smoke.test_project_source_index covers provider-none/off with no external construction, required versus preferred-unavailable behavior, executable minimum version, malformed/stale status, source/database changes during observation, bounded output and timeout, and a real provider sync/check when CodeGraph is installed. tests.smoke.test_source_to_specification covers source-bound observations, tampered skill identity, unresolved behavior blocking promotion, explicit review and selective refresh. These are boundary/custody checks, not semantic-equivalence proofs.

Held-out impact tasks should include a public symbol rename with a dynamic caller, a shared configuration contract change and a dependency version change. Held-out inverse tasks should include contradictory tests, undocumented error behavior and two source revisions with one changed boundary. Review expected affected tests and source/specification behavior independently before model exposure. Measure impact recall and false positives alongside held-out regression success, repair cost and human interventions; count unsupported results and abstentions. No live comparative trial is reported in this checkpoint.

PR 369 review checkpoint

Reviewed revision e2f62cc996760505babee8fa3f20260327fbca12 of existing PR #369; do not merge it solely because CI is green. Two focused disposable-directory reproductions confirmed defects in its new native-project registry:

• P1 — ancestor symlink escapes ownership. _native_project_roots checks the final child for a symlink but not its parent chain or resolved containment. Pointing .literate/native-projects at an external directory makes a correctly pinned child there pass registry validation. Require component-by-component safe paths and containment on both read and registration; test symlinks at both parent levels. Exact reader.
• P1 — concurrent registration loses a successful child. The read/append/replace sequence has no shared mutation lock or compare-and-swap. Synchronizing two writers immediately before replacement makes both return success but retains only one registration. Hold the existing project mutation boundary across validation, child installation and registry update; revalidate stage and registry under that lock, and prove rollback does not delete another writer's child. Exact writer.

Both findings belong to ADOPT-PROMOTION-001, not a competing implementation. They were reproduced against the downloaded exact-revision helper, not asserted from a different branch's tests. PR #369 subsequently merged upstream as e1e602156548cbb8ac12cfadc1b76f16ee48590b; this work did not merge it or post a remote comment. With explicit operator acknowledgement, the local branch integrated that merge at 96afb9d4, preserving the audit checkpoint 437e1127. Both registry defects are repaired locally. Forty-seven promotion, registry, lifecycle-lock and onboarding tests pass, including rollback preserving a foreign child and stage mutations sharing the lifecycle lock. A separate fresh converted-project journey also passes: actual retained-harness execution, receipt promotion, retained-stage advancement, signed static integration and drafted advancement, with original source bytes preserved. Combined-revision broad CI remains pending. PR #364 is a separate planning correction and says P0.2 is already assigned; reconcile the local architecture draft with that work before landing it.

Implementation checkpoint

Retained-scope refresh (#368) is implemented locally with exact acknowledged plans, policy rebinding, preserved conversion history and receipt invalidation. Eleven focused refresh tests and the retained-harness/adoption/performance regression batch pass (61 distinct tests). Combined-revision full qualification and landing remain pending. The peer survey also found library-artifact PR #370 merged upstream and issue #280 closed; review and reconcile that implementation before duplicating library work.

Multi-root adoption (#365) now exposes candidate roots and accepts an explicit --root-plan with exclusive ownership, declared shared consumers and per-root commands. Exact raw source bytes and executable bits bind the reviewed plan. Nineteen planner/refusal tests pass, including races and stale-plan rejection; the wider adoption/harness/Git-boundary batch passes too (106 distinct tests). Formatting and lint pass on 865 files. Refined apply intentionally refuses until transactional per-Component materialization, independent receipts, refresh and boundary transfer are implemented. The default single wrapper is unchanged. This is a completed planning checkpoint, not completion of #365 or release readiness.

The next internal #365 checkpoint stages distinct source-free Component projections and earns component-scoped receipts through actual disposable retained-harness execution. Its 17 new tests cover positive counts, private/shared invalidation, scope drift, tampering, acknowledgement, contention and rollback. The combined adoption/retained-receipt batch passes 78 tests. Public refined apply remains disabled: conversion, project-wide verification, refresh and boundary transfer must consume this adapter before the end-to-end adoption gate can pass. No source was moved or retired by this implementation checkpoint. The current peer survey also found #371, now recorded as a library/promotion follow-up to review alongside merged PR #370.

Seven bounded audit items are closed against focused evidence: AUD-REL-001, AUD-UX-001, AUD-MCP-001, AUD-SKILL-001, AUD-CI-001, AUD-INTEL-001 and AUD-RESEARCH-001. This does not close the release umbrella or establish efficacy. The merged 1.1 projection contains 21 explicitly targeted items, eight closed (including upstream DRAFT-PROMOTION-001). ADOPT-PROMOTION-001 remains open for the audit's additional qualification; upstream closure does not satisfy that work.

Validation completed on this development tree: 165 focused lifecycle/onboarding/ MCP/intelligence/schema tests; 20 packaging/research/roadmap/skill/pin tests; 15 isolation-policy tests; the self-describing project regression; formatting, lint, changed-skill evaluation, 260 Mermaid diagram renders, two strict OpenSpec changes, and the current driver review. Codex's plugin-creator manifest validator also passed on the development archive, not a released plugin session.

The first broad Python attempt stopped after source changed during its self-replication fixture; it is invalid qualification evidence. The stable-source rerun stopped at 2,587 tests: 2,560 passed, 26 skipped and one failed; 1,070 selected tests were not reached. The failure was a CI assertion still counting actions/checkout@v6 after immutable pinning. It now enforces full history per checkout step, and the regression plus three pin-policy tests pass. The broad gate must still rerun; neither this repair nor the interrupted run establishes release qualification. litai verify . passes authority, skips locks and provider-none intelligence, and fails the stale project test receipt. The live contribution sweep remains not ready: the previous sweep found nine issues and two reviews without current-release dispositions. The post-approval peer survey now finds eight open issues and one open, green review (#364); release dispositions still need a fresh sweep. The audit checkpoint and main integration are committed locally; no final artifact, installed-upgrade, hosted CI or terminal-collateral claim is made for this batch. No tag, release, PR merge or external publication occurred.

Next checkpoint: ADR 0039 is accepted. Preserve the existing accepted HTML contract, continue multi-root adoption from the locally tested retained-refresh implementation and exact-revision qualification. Main integration is complete; no included feature is silently deferred.

Release boundaries

The current release scope includes the audit fixes and existing 1.1 commitments; no existing item is silently deferred. Production containment requires a product-scope decision grounded in the existing threat model: a locally authorized development release may truthfully say it is uncontained, while a production-isolation claim requires qualified backends and authenticated evidence. Open experiments remain unmeasured; implementing evaluation tooling does not establish research superiority. Preserve historical roadmap entries until their individual closure evidence is verified.

Release preparation ends before tag push and public publication. Current policy requires terminal document collateral and a clean release line; surface any missing authority or credential as a concrete blocker after all independent local work is complete.

Review-process coverage reconciliation

Observed 2026-09-16 UTC against roadmap branch work/1.1-roadmap-diagnostics following 09d3c7da. PR #438 selects 81fbee6a99986d149a53d2a3d85360965f5e495a. This is a coverage audit, not release approval or a claim that unchecked work passed. RELEASE-REVIEW-001 retains each terminal review gate; RELEASE-INTEGRATION-001 owns consolidation and tracker dispositions. Prior tracker counts in this report are dated history. All current open issues and reviews now have explicit roadmap ownership.

The score-improvement program has 21 milestones with unfinished tasks; all have same-ID owners in the active queue. Its 112 unchecked tasks remain required, including longitudinal evidence where named. The defect-mitigation program keeps remaining Phase 2 shape enforcement, accepted ADR 0011 implementation and Phase 5 reference-sample proof under MITIGATION-FW-001. The SDLC audit's ordered finding matrix above retains its original owner and acceptance for each finding. No parent is closed by this inventory.

Coverage recheck after qualification repairs

A fresh 2026-09-16 UTC tracker inspection at 406b53d1 confirms 17 open issues and 28 open reviews. Every issue and review has a cited queue owner; every work ID in this audit resolves to an active-queue entry. All 21 unfinished score-program milestones have same-ID owners, retaining their 112 unchecked tasks. GitHub's full open-review thread inventory reports zero inline threads, with no omitted PR or thread pages. This does not replace the independent code/security review gates or classify unmerged work as accepted.

The strict 1.1.0 --require-all-open-targeted projection succeeds with 117 open items and ready: false. All review-process obligations in the coverage table remain roadmapped. Current work adds SDK qualification performance and loopback probe repairs under RELEASE-INTEGRATION-001. The maintainer explicitly accepted ADR 0043 for 1.1 on 2026-09-16 UTC; worker-health implementation and qualification remain open under WORKER-HEALTH-001. Live cleanup still needs exact-target authority.

The final scope-aligned contribution sweep is sha256:02031457dd60dd8dc75eb8c8ba20590ad803c2271ddc0960b2b8bc52c4d3f429. All 17 issues and 28 reviews now have matching 1.1 milestones and include dispositions. This corrects nine issue deferrals (#251, #295, #296, #365, #379, #380, #403, #411, #440) and PR #437's deferral to 1.2, which conflicted with the maintainer's full-integration directive and their explicit queue targets. PR #441 also has its current include disposition. The sweep reports 45 included open tracker records and 74 unclassified branch/worktree records (37 of each), with readiness false. RELEASE-INTEGRATION-001 owns their exact inclusion/disposition and landing evidence; mapped ownership does not close release obligations.

Current issue ownership

All 28 current open PRs are mapped individually in the active queue's current review inventory, including the native SDK stack, C++ library, Conan verification, worker DAG, CI updates and consolidated release PR. The coverage pass discovered #411's missing native npm package owner and added NPM-PACKAGE-001 with public plan/build/verify and negative archive-custody acceptance. Worker health issue #440 remains WORKER-HEALTH-001.

The strict release-scope projection also found ten existing open items with no explicit release target: RELEASE-CANCEL-001, UPDATE-CLOSURE-003, UPDATE-FOLLOW-002, ACCEPT-JSON-001, NATIVE-SDK-001, CPP-LIBRARY-001, CONAN-VERIFY-001, MARKDOWN-432, RETAINED-RECEIPT-002 and WORKER-DAG-001. All now explicitly target 1.1.0. The two CPP-LIBRARY-001 sections are consolidated with all implementation and qualification obligations preserved. Including NPM-PACKAGE-001, the corrected projection contains 118 open items at reconciliation, reduced to 117 after the real nested workflow sample qualified and WORKFLOW-CATALOG-001 closed. The earlier 107 count omitted the ten untargeted items and predated the npm intake. summarize_release_roadmap.py 1.1.0 --require-all-open-targeted succeeds and still reports release readiness false.

Contribution sweep sha256:2cefc39a42892daf684e5210b71d0e7c7d2c03d78383d4555e637572af99001d reports 74 unclassified records and 34 included open tracker records; readiness is false. Mapping into this roadmap does not alter forge milestones or close issues. Repeat the sweep after integration and resolve its remaining records before planning publication.

Immediate qualification follow-up

The preserved c7dac9a19 release run stopped after 1,160 tests with one missing-Conan failure and 22 skips. The repository-pinned developer tools have been installed and the checkpoint repair resumed. A resumed repair cannot supply release attestation, and an older candidate cannot qualify PR #438's newer head. Repair the clean-session bootstrap prerequisite, preserve the failed observation, then require a fresh full run and current installed/live/platform evidence on the final integrated source.