Ecosystem planning review — 2026-09-28

Tracked by pm-p9a4, pm-m08hza, pm-1jzupz, and pm-wg07.

Project management = context management. This review records a complete structural census, targeted semantic review, measured operational checks, and planning decisions on existing canonical items. It does not certify that every historical sentence is correct or that every future capability is implemented. Read the live item before acting on this dated evidence.

Population and evidence

The strict CLI snapshot at 2026-09-28T05:40:12.139Z contains 2,826 items: 2,532 closed, 221 open, 59 draft, and 14 canceled. No records or fields were omitted, and no item or directory was unreadable. Blocking is also a derived dependency condition: the initial orientation reported 60 blocked open items, although no item used a literal blocked lifecycle status.

The cursor-resumed event read collected 68,876 complete events from 2,841 streams through 2026-09-28T05:41:59.143Z, including deleted-item and workspace history. All 23 Plan items also received deep plan reads for steps, decisions, validation and resume metadata. This later cutoff includes the first review claim/comment events; it is not silently equated with the earlier item snapshot. History-drift validation passed across all 2,826 current items. The public SDK independently verified hash chains for all 2,841 collected streams with zero failures.

The per-item census records every snapshot ID, status, content fingerprint, structural classification, missing planning fields, historical evidence gaps, and remediation owner. These are mechanical findings, not invented historical estimates, outcomes, or review approvals. The full raw corpus stays outside public documentation.

Measure	Initial observation	Interpretation
Directed graph edges	14,097	Deduplicated directed basis
Ordering layers / critical-path edges	12 / 11	Acyclic prerequisite graph
Missing graph endpoints / isolated items	0 / 0	Across all statuses
Active / terminal outcome-unreachable items	0 / 0	Typed reachability basis
Generic `related` edges	5,589	39.65% of directed edges
Specialized semantic edges	1,885	Discovery, recurrence, incident, verification, supersession
Items without a specialized semantic edge	1,463	86 active, 1,377 terminal; not necessarily invalid
Legacy hierarchy exceptions	7 children	Existing informational cardinality/direction findings
Closed items missing structured resolution evidence	170	Preserved residual, not a new regression
Active items missing required planning fields	1	Sentry major-upgrade item, repaired in this review
Tracked non-tracker artifacts with exact item links	1,921 / 1,921	Path ownership, not proof for every code change
Duplicate candidate clusters	28	Reviewed candidates, not 28 duplicate obligations

Scanner records with different provider identities, parent/child acceptance slices, distinct incidents and dated release executions remain separate. The three identically titled accidental probes are already canceled. Similar audit titles require their own event/acceptance comparison before any consolidation. No existing item was deleted or closed merely because it matched a title.

Graph organization

Use one structural parent and outcome paths for scope, implements for delivery, blocked_by for an actual prerequisite, discovered_from for provenance, verifies for verification responsibility, and recurs_from only for an established recurring mechanism. Maintain prior evidence when correcting a link. A declaration that an item will verify something is not a passing test result.

The native reconstruction plan retains its original execution obligations; overlapping newly added steps are explicitly superseded by their existing owners, and a dated census checkpoint records the completed refresh.

The planning pass restores explicit missing prerequisites in prospective reminders, cross-branch relevance and external-package verification. The historical batch restores three discovered_from edges to the governance run whose original comment explicitly named the discoveries. Their release and closure evidence remain unchanged. Generic links and transitive ordering shortcuts are not added to meet density or depth targets.

pm-qnl2 owns active enrichment; pm-lnrk owns historical batches; pm-pj2isb owns misleading placeholder metadata. The completed pm-e9zh preserves its explicitly unrecoverable 170-item residual. A missing historical observation must remain unknown until independent evidence is found.

After this pass the live graph contains 2,829 items and 14,118 directed edges, including 1,894 specialized semantic edges. Generic related is 5,590 after one reviewed correction from causal provenance to non-causal coordination; no dangling endpoints, isolated items or ordering contradictions were introduced. Outcome reachability remains 100%. The 1,458 records without a specialized semantic edge remain visible for evidence review; absence alone does not justify inventing a relationship.

Horizon and outcome organization

These are planning windows, not delivery promises. Capacity, prerequisites and measured evidence govern promotion. Priority measures impact; horizon measures readiness and uncertainty. Review monthly and when a material observation changes.

Horizon	Illustrative window	Canonical outcomes and promotion evidence
Current	0–3 months	Agent-legible grammar, active readiness, retrieval failures, release diagnosis and current protocol proof. Require completed-task token comparisons, preserved compatibility and actual installed-consumer evidence.
Emerging	3–12 months	Provable history, reproducible observations and measured self-improvement. Require independent replay, mutation coverage and three consecutive measured improvement cycles.
Expansion	12–36 months	Proven scale, universal domain platform, federated agents. Require published million-item/fleet distributions and SDK-only domain exemplars.
Exploratory	36+ months	Multi-decade continuity and independently evaluated cognitive mechanisms. Require explicit architecture decisions, migration proofs, calibrated uncertainty and rejection criteria.

The SDK-complete outcome and GA stability contract retain their own acceptance boundaries. Earlier SDK promotion work remains completed history; future integrations must use published SDK contracts and expose gaps as individual items. New names or parallel roadmap trees do not improve traceability.

Command simplicity: root help already has thirteen entries including help. Evaluate the remaining grammar through complete orient/retrieve/claim/mutate/ verify/close and recovery journeys. Preserve working aliases and capabilities. Count tokens across errors, retries, cursor pages and contract discovery. The new Plan receipt issue records whole-plan output amplification observed during single-step mutations; its acceptance requires bounded SDK-owned receipts and preserved full inspection.
Algorithms: the existing portfolio remains authoritative. Individual items now detail bounded candidate quotas, evidence coverage per token, typed spreading activation, activation decay, approximate affinity sketches and content-addressed summary trees. Each names inputs, costs, fallback, negative controls and a promotion/rejection rule. Mathematical guarantees require their assumptions; a heuristic is not a proof.
Scale and branches: separately measure cold build and warm reads at 3, 100,000 and 1,000,000 items, and 1/10/100/1,000 agents. Exercise partitions, retries, expired claims, clock skew and arbitrary merge order. Record p50/p95/ p99, memory, index amplification and no-lost-acknowledged-write evidence.
Recursive improvement: bind observe, propose, evaluate, land and remeasure stages to immutable evidence. Keep the evaluator and held-out data independent of the change under evaluation. Rejected proposals stay searchable.
Universal environments: episodes use SDK reset/observe/step/score contracts, seeded snapshots, idempotent actions, independent verdicts and temporal cutoffs. Distinguish termination from truncation and prevent future-evidence leakage.
Durability: append-only API behavior, hash-chain tamper evidence, authenticated authorship and externally anchored proof are different guarantees. Migration, redaction and cryptographic rotation need explicit restore tests.

Verified behavior and remaining limits

Local qwen3 retrieval

The project already selects local Ollama qwen3-embedding:0.6b. Refresh embedded 1,085 stale inputs in 34 requests, with zero retries or splits, at approximately 11.91 inputs/second. This proves multi-input batching at the current corpus size; it does not prove the outstanding 100k/1M performance acceptance.

Five project golden queries were evaluated at k=10:

Mode	Mean nDCG	Mean MRR	Mean recall	Broad history/episode query recall
Keyword	0.7451	0.80	0.8727	4/11
Semantic	0.7944	0.85	0.8182	1/11
Hybrid	0.8405	0.90	0.8364	2/11

Aggregate passes conceal weak queries. The existing retrieval-quality owner pm-p1g43q now carries this observation, feasible per-query floors, held-out paraphrase evaluation and mode comparisons. At k=10, a query with eleven relevant items cannot attain recall 1. The final semantic index check covered all 2,829 current items with zero stale vectors.

MCP and SDK

The official versioning page identifies 2026-07-28 as current. The repository conformance matrix was compared with the official key changes. The focused SDK, stateless stdio, Streamable HTTP and legacy-handshake run passed 145 tests in twelve files. Exact-version public verification also passed direct and package-selected npx/bunx CLI dispatch, MCP discovery, Skills, Apps and HTTP, including missing-bin negative controls. This is focused conformance evidence, not an assertion that every optional extension or every deployment is certified. Deprecated features must not be newly advertised merely to increase a feature count.

Ordinary review mutations recorded detected Codex identity, model and effort without PM_AUTHOR or author flags. Role/topic inference carries its source and confidence. Missing signals remain unknown; configuration is not native proof.

Daily release

Version 2026.9.28 exists in npm and GitHub. The original automatic attempt failed after Windows global npm installation failed in the child release run. The same-version recovery passed npm, global npm and Bun on Linux, macOS and Windows and advertised the existing version. It did not create a second automatic version that day.

The new installer-diagnostics issue records why the original cause remains unknown: the subprocess wrapper discards error/signal details and the installer labels every retry registry availability. Recovery proves the artifact can pass; it does not erase the failed attempt or establish its cause. Keep automatic reliability, publication and consumer acceptance as separate measurements.

Prevention and repeatability

Six active-only zero-debt assertions now join tracker-context-quality for missing description, acceptance criteria, risk, confidence, estimate and expected result. The existing local and hosted quality pipeline invokes this same native SDK gate. Terminal evidence gaps are deliberately outside the active-only policy. An isolated temporary workspace failed all six assertions with missing fields and passed all six after repair. This checks presence; semantic quality still requires evidence review.

For repeat reviews, follow the ecosystem planning skill. Use complete CLI snapshots outside the prompt, cursor-resumed event streams, exact source fingerprints, all-status duplicate checks and bounded mutation batches. At million-item scale, use streaming/indexed consumers rather than loading a whole JSON collection into the agent's context. Every changed item retains its canonical owner, history and original acceptance boundary.

Verification record

Quiescent validation passed for all 2,829 final items. The 170 historical structured-evidence warnings are unchanged. All 38 changed history streams pass the public SDK chain verifier; the 35 pre-existing streams retain byte-identical historical prefixes. The three enriched closed features retain identical status, closure, resolution and release fields. Final graph-composition and all 47 tracker-context-quality assertions pass, including the six new readiness rules.

An earlier long-running context gate saw one history-drift failure while this review was still mutating the tracker. Independent validation and the final gate pass with writes stopped; the earlier result is retained as a concurrent-read observation, not rewritten as success. Skill/document validation, local links and the generated changelog check also pass. All review claims are released. Three distinct issues were added after duplicate checks: installer diagnostics, Plan mutation receipts and startup admission diagnosis. Existing programmes retain their original acceptance boundaries.

The broad static run passed its build, lint, generated contracts, SDK parity, command grammar, task-token, MCP deprecation, gate-registry and context-quality checks, then failed the import-cost timing gate (bare Node median 305 ms against 106 ms). An isolated rerun passed all ten SDK entrypoints with unchanged budgets. The initial timing failure remains part of the evidence; it is not a continuous all-green static invocation.

The remaining graph and record-integrity gates passed. The mutation gate passed with 323 of 330 mutants killed and seven classified as equivalent. The CLI transport benchmark initially exceeded its list threshold (442 ms versus 419 ms); the isolated rerun also failed: get 385/362 ms, context 466/433 ms, next 457/440 ms and create 424/398 ms. No performance limit was raised. The new startup investigation distinguishes runner qualification from a demonstrated product regression and preserves the completed transport-floor implementation. This is an unresolved validation limitation.

PR review corrections

PR #1333 identified two overly broad relationship claims. The standing release reminder predates the new installer-diagnostics issue, so their link is related coordination rather than discovered_from provenance. Four verifies rows from the historical batch were removed: checking citations, history continuity and release attribution does not verify the source run or feature behavior. The three source-backed discovery links on closed features remain valid and unchanged.

The original writes and their corrections remain in immutable history. Five holder/target/text-scoped citation exemptions distinguish the batch inventory from verification claims without raising the prose-edge ceiling. The corrected graph-composition gate passes.

The follow-up review narrowed those five exemptions to the complete original paragraphs, including their historical source-event coordinates. In an isolated temporary tracker the original citations pass, while a later batch comment asserting an unrecorded prerequisite is blocked with one detected gap. This prevents the common phrase “Historical batch” from exempting future claims.

Annotation review corrected 26 notes and five comments that contained literal JSON wrappers around prose. Supported note/comment edits retain the exact inner text and append correction events. Per-item rereads confirm lifecycle, closure, release, expected and actual fields are unchanged; original annotations remain in the history streams. Planning guidance now distinguishes plain-text input from intentional JSON events.