Skip to content

Current milestone

  • Independent reliability repair — ADR-0205: The owner authorized safe failure diagnostics, bounded recovery, cleanup isolation and scoped reliability reporting on codex/app-reliability, with its own pull request to main. This repairs existing app behavior without importing the parallel memory retrieval workstream or changing its activation gates. Milestone 15 remains incomplete; monitoring activation and production delivery remain separate steps.

  • Milestone 33 — General memory search and recall: the owner authorized the six-stage plan on 2026-10-11 under ADR-0178. The checkpoint comparison completed all 1,600 observations on 2026-10-19 and failed the original recall, invalid-current-fact and per-case cost thresholds. The owner approved the separate bounded rollout in ADR-0220, subject to actual human findings, production compatibility, final review/CI and verified rollback. This authorization does not mark M33 complete or claim deployment readiness. General memory recall registers twelve gates for cross-category evaluation, relevance before candidate limits, measured semantic retrieval, task-aware queries, governed source recovery and release evidence. The category-independent candidate repair is implemented; the 400-seed diagnostic improves recall from 67% to 68.75%, leaving semantic/implicit-use work open. The synthetic semantic comparison reaches 97% top-20 candidate recall; the local pgvector index and shared experimental recall path are implemented with source-version fences (ADR-0179). Bounded provider admission and persistent indexing slices are implemented (ADR-0180). Explicit query preparation now feeds the local cache under the same durable budget (ADR-0181). Run-bound shared recall now binds saved task state and effective query policy (ADR-0182). Context assembly now re-reads current evidence instead of retaining completed recall (ADR-0183). Release-bundle validation now recomputes cross-category recall and implicit answer lift with exact-version admission and revocation (ADR-0185). Source indexing now has default-off configuration and maintenance composition, bound to independently reviewed release dependencies (ADR-0187). A verified offline CPU adapter now supports restricted inputs without remote disclosure, and local run authority plus explicit preparation are composed on both recall paths (ADR-0188). Governed source recovery now has shared correction fences, source-copy erasure, bounded rendering and default-off composition (ADR-0189). Optional query preparation now reserves a durable run budget before provider work (ADR-0190). A separately admitted relevance reranker has semantic, lexical and linked-record regressions (ADR-0191). The development harness now exercises actual runtime context and tool results with gold-blind inputs and exact-value scoring; source recovery remains distinct from atomic formation (ADR-0192). Bounded deterministic task-context composition is implemented as a default-off ablation (ADR-0193). The PostgreSQL harness now uses actual elapsed worker time, normal People formation and verified paired formation-response banks under one experiment budget (ADR-0194). Sparse development controls complete twenty seeded and twenty formation observations; they expose retrieval and labeling failures and do not pass the release gates. The owner approved the 400-scenario corpus on 2026-10-13; explicit factual questions are separated from implicit tasks, with incident and per-family/split coverage checked. The paired pilot exposed a formation@11 defect: named-person statements were rejected by the older owner-subject rule, and separate People claims could merge under one subject. Source-validated named subjects and whole-statement comparison now have regressions without changing formation@9 or People identity admission. Shared rendered recall now caps retained deliberate results plus automatic base/delta recall at twenty items and two thousand tokens per request, with current-source validation and mandatory corrections (ADR-0195). Actual M33 searches now consume the durable run allowance, including sparse fallback, automatic deltas and People discovery (ADR-0196). Exhaustion returns a limitation rather than a false empty search. The paired PostgreSQL pilot completes all five arms: sparse 8/20, hybrid 6/20, reranking 18/20, source recovery 6/20 and combined 18/20 correct answers. This development comparison predates the named-People repair and final budget enforcement; it is not activation evidence. Offline publication now requires complete original runtime measurements, settled provider journals, matching formation banks and independently supplied reviews. It recomputes the canonical observations and refuses source or cost discrepancies; actual confirmatory dependency binding and reviewed release inputs remain open. Current People query narrowing now preserves prepared encoder input and all hard restrictions, with identity filtering before sparse and dense limits (ADR-0197). Its shared contracts and PostgreSQL checks pass; comparative release evidence remains pending. Live development failures remain retained, including People-path ranking and incomplete formation. Historical semantic revisions now have shared visibility, invalidation and dated People contracts plus an actual PostgreSQL tool/reader-context regression (ADR-0200). Reviewed held-out formation-to-answer evidence remains open. Runtime assembly now records one bounded combined ordering for scoring, with unknown or truncated evidence refusing certification. The six-arm task-context pilot completes 120 observations: reader-planned search answers 16/20 versus reranking with a scripted query at 10/20. All four remaining reader-planned failures exclude correctly formed claims through guessed hard type filters. The model-facing tool now uses soft type hints (ADR-0198), with legacy explicit filters preserved. The frozen pilot predates that repair and exceeds the exploratory reader-latency ceiling in two arms. A separate PostgreSQL load measurement retains 240 retrieval samples across 100, 1,000, 10,000 and 100,000 synthetic distractors. Every sample ranks and renders the older stronger match; fixed vectors exclude encoder/model costs. This measures database/retriever load, not end-to-end release latency. The owner approved verified-input release binding (ADR-0199). Its implementation preserves original observations and rejects changed evaluated inputs across a merge. The confirmatory runner now captures deployment composition, verified model assets, actual PostgreSQL schema and frozen controls before held-out task parsing. Production startup binds the same schema fingerprint. Per-observation confirmation now binds all twelve captured runtime variants and independent principal attribution before model work. Confirmation preserves the production model pin and chat reasoning settings and verifies the actual checkpoint's reader digest; the earlier diagnostic reader mode cannot supply confirmation. Collection and publication reject unbound or retagged development rows. The owner authorized one USD 100 aggregate cap for the 4,800-observation confirmation; one journal covers every variant and interrupted attempts retain their rows and uncertain reservations. The owner approved the exact 400-scenario corpus on 2026-10-13; its retained independence receipt binds the original corpus and all scenario identities. The production-reader sparse baseline completed twenty development observations with six correct answers; it does not certify quality. Native historical-index load measurement completed 480 samples through 100,000 competing records, with every target ranked first and rendered. The first confirmation stopped at 34 of 4,800 observations because a required test-fixture correction changes its bound inputs. Original records and failed status are retained; aggregate reported and held cost is USD 5.45365250. The owner subsequently approved a USD 300 aggregate cap and one replacement after final review and CI; its USD 290 allocation includes no forgiveness of earlier charges. Foundation PR #159 deployed successfully as 81062c5b; PR #158 retains review, exact-head CI and deployment requirements. The replacement on reviewed/verified 7214a8df failed at 4 of 4,800 observations: an unused automatic preparation reservation expired before a delayed deliberate search. The evaluator now isolates deliberate preparation as well as context; production budget enforcement is unchanged. Originals remain retained under 20261014-confirmation-v2/, with aggregate reported and held cost USD 5.62063000. V3 subsequently failed at 1,217 observations on 74b14815; its original archive and scratch database remain retained. A mandatory incident returned clarification despite rendered evidence, and a diagnostic defect obscured the terminal provider failure. Aggregate reported and held cost is USD 55.913546. The owner requested completion before merging PR #163 on 2026-10-15. ADR-0206 records the scope and diagnostic repair; a new complete run may use at most USD 240 after ledger reconciliation, CodeRabbit review and exact-head CI/signing. Old measurements cannot certify the changed evaluated inputs. The approved recovery now isolates invalid interaction proposals from grounded claims, preserves named-subject corrections, discovers unambiguous people through governed relationship evidence and shares their context with planned recall. Retained evidence validates its identity and temporal dependencies, and People retrieval preserves prepared semantic authority (ADR-0208). Two isolated evaluation workers share one allowance without changing the observation matrix or release thresholds (ADR-0207). Focused regressions pass; combined verification, review and exact-head CI must precede new paid collection. Reviewed reliability and the independent dreaming release are integrated for existing PR #163, with migration ancestry ea3310a1b003 → ea3210a1b00d → ea3310a1b004 → ea3310a1b005. The later v5 on f837ed3c stopped at 124 observations; its diagnostic identifies a provider-key-shaped match in OpenAI encrypted continuation, whose original meaning remains unknown. The owner approved the narrow ADR-0210 repair on 2026-10-16. Standalone secrets and other providers/fields retain rejection; original replay bytes stay unchanged. Failed attempts remain immutable. Aggregate reported and held cost is USD 88.371792 under the unchanged USD 300 cap. Source verification and synthetic native load do not establish full quality; the repaired source needs fresh review, ordinary CI/signing and full evidence. ADR-0216 separately authorizes ordinary signing on the final reviewed reliability commit with a temporary exact-revision context restriction, preserving deferred dev work and restoring the original restriction afterward. Complete confirmation, independent failure review and release publication remain outstanding. The later v6 on 16153a5f stopped at 1,239 observations after a cancelled formation call. All originals remain immutable; eight ledgers total USD 135.90882425 including uncertain reservations, leaving USD 164.09117575. The owner approved checkpointed continuation and a focused production comparison (ADR-0212): all 400 scenarios, both paths, formation sparse control versus final hybrid, or 1,600 required observations. Reuse requires unchanged runtime inputs, verified original rows and banks, and complete accounting. The separately verified collector is collecting the remaining positions with append-only checkpoints. The source fixes and current production changes are being integrated on codex/recall-production-release; complete scoring, actual independent human failure review and deployment compatibility remain open. No quality threshold or independent human review is waived. Work proceeds independently on codex/memory-retrieval-evaluation; unrelated development and releases are not dependencies or held by this run. The owner explicitly authorized the recall pull request when ready, its review, hosted verification and production deployment. The separate hold on pushes to dev remains active. Optional-channel activation remains evidence-gated. M16/M32 frozen controls and unattended dreaming requirements remain unchanged.

  • Milestone 32 — Memory reconsolidation (dreaming): authorized on 2026-10-02 as a parallel workstream. memory-reconsolidation.md and ADR-0169 define twenty-four registered gates. The owner authorized ADR-0174 on 2026-10-10: daily pending proposals, a private web review page, and explicit individual approval applying changes to real memory. This supervised path is independent of the unchanged unattended quality gate; no further blinded review packet is required for this path. The implementation now includes bounded durable inventory and provider execution, lossless equivalence merging, extractive summaries, independently supported tentative connections, conflict history, source invalidation and erasure, and HTTP/CLI/native owner controls. Both stores share the operation contracts. Connections use the existing derived memory path and retain every original; they remain inferred at confidence at most 0.35 and expire no later than thirty days after their latest original evidence. The native Synthesis journal exposes summaries, connections, conflicts, supporting originals and previewed merge undo. Background maintenance now composes one bounded proposal/verification slice behind default-off processing and exact release evidence. Removing the evidence stops new work and derived recall while preserving history. The frozen offline control still measures retrieval only. A separate live command freezes answer-bearing original-only observations before measuring merge-only and connection-enabled arms, with three repeats, fixed budgets and all failures retained. Publication recomputes the observations and requires matching upstream evidence. Two earlier comparisons retain their failure evidence. The owner approved ADR-0170 on 2026-10-08. Its versioned offline review contract exports blinded packets, requires complete human decisions and recomputes scores. A third, contract-bound comparison preserved all facts and had zero false merges, but still failed answer lift and development merge-answer nonregression. The owner judged all fifteen candidates: ten equivalent and five not equivalent. Recomputed precision is 75% on development and 57.14% on holdout, below the unchanged 80% floor; recall passes both splits. Equivalence review cannot excuse answer failures. The October 9 repair rejects hedged original copies and duplicate-claim connections, distinguishes summaries in the proposal/verifier instructions, and preserves the best visible member's recall rank after merging. Its new comparison records 14/18/14 development and 17/17/19 holdout answers out of 36. Merge-answer nonregression passes, but 0/5.56-point answer lift still fails; eleven new hypotheses remain unverified. The owner has stopped repeated review requests; the completed October 8 review and real-data usefulness review stay complete. The optional PostgreSQL summary-reader repair followed that run, so its diagnostic recording is not evidence for the final source tuple. The owner approved ADR-0171 on 2026-10-08 for a separate, isolated evaluation using newly confirmed fixture text from their own memories. The isolated runner, durable experiment budget and local input/output review are implemented. The confirmed four-input experiment completed and the owner judged its accepted summary useful on 2026-10-09. That qualitative review is finished; it supplies no comparative or activation evidence. Any new experiment still requires exact-text confirmation naming the provider/model and USD 0.25 ceiling. Production attribution is unchanged. Owner-reviewed production mode is live under ADR-0174 after authorized pull requests 156 and 155, clean final-code review, pipeline 1066 verification (including macOS and iPhone/iPad journeys), and pipeline 1067 delivery of d053d1d9. The API, documentation and website identify that release; TestFlight upload passed. Daily proposals are enabled in all application roles and only individual owner approval applies them to memory. The first automatic batch deferred with no proposals or provider calls; its next claim is due October 12 at 01:30 UTC. The second run also deferred without provider calls. The October 11 diagnosis found legacy erased People heads with missing reverse attribution; their existing source refusal remains intact. ADR-0186 implements persistent review sign-in and explicit blocker/deferral reporting, pending delivery. Authorized metadata-only recovery found no complete surviving identifiers. On October 13 the owner authorized forward-only dreaming (ADR-0202): a durable cutover excludes existing memories and source events while allowing new evidence through existing safety checks. Focused regression coverage passes; production cutover awaits verified delivery. Unattended application stays disabled. On October 15 the owner authorized an independent dreaming release and its dedicated pull request; supervised delivery does not depend on Milestone 33's remaining recall evaluations. Migration ea3210a1b00d follows production ea3310a1b003, with later recall migrations rebased before their own release. That release subsequently completed through PR #165 at 0c7b0d74: all ten hosted checks and every delivery job in pipeline 1141 passed, including TestFlight. The public API, documentation and website identify 20261015-185949-0c7b0d7. The forward cutover activated at 2026-10-15T19:05:44.185661Z with no blockers, preserving all 469 unknown erased heads, five historical run receipts and the existing daily schedule. HTTPS review-only sign-in works with thirty-day cookie persistence. Supervised review remains enabled and unattended application remains off; this operational delivery does not supply the missing comparative quality or independent review. The once/dry-run operator controls, remaining complete runtime gates and fresh matching comparative release evidence remain required; no verified sequential ceiling advances.

  • Active milestone: Milestone 15 — operational hardening — is the next sequential authorized milestone, is specified by operational-hardening.md with sixteen gates, and has not started; its backup tranche has no dependency on the milestones before it. Milestone 14 — inbound surfaces and pairing, specified by inbound-surfaces.md with twenty-one gates — completed on 2026-10-07. A paired message's run has its own USD 10 budget (ADR-0064 amendment); the sixteen findings pull request 104 left open were resolved and pull request 143 carried them to main; the owner's production smoke with the Telegram surface runbook passed on 2026-10-06; and pull request 153 shipped the smoke's one finding with hosted CI and a clean CodeRabbit review. The completion was recorded on 2026-10-09. Milestone 13 — general-purpose subagents and delegation, specified by subagents-and-delegation.md with twenty-one gates — completed on 2026-08-27 when pull request 72 passed hosted CI and a clean CodeRabbit review on its final head; the completion was recorded on 2026-09-28. Milestone 17 — the memory read API and the native memory browser — is specified by memory-read-api-and-browser.md with ten gates and completed as a second parallel workstream on 2026-08-24 without advancing the verified sequential ceiling. Milestone 18 — first-class email integration — is specified by email-integration.md with seventeen gates and is a third parallel workstream on the same terms, its original shared-file touches named in ADR-0071 and its operator-managed multi-account extension in ADR-0085. Milestone 19 — conversational schedule creation — is specified by scheduling.md with five gates and was a fourth parallel workstream, limited to one-time creation through the model; it is complete. Milestone 20 — calendar recurrence and conversational schedules — is specified by scheduling.md with six gates and was a fifth parallel workstream, adding monthly and yearly rules and conversational creation for all four recurring calendar kinds; it is complete. Milestone 21 — adaptive memory distillation — is specified by adaptive-memory-distillation.md with thirty-one gates and is a sixth parallel workstream, adding integrated episodes, high-recall direct and hypothesis formation, and evidence-based forgetting; ADR-0090 widens its existing source-grounding gate to attributed communication without changing the gate count. Milestone 22 — the persona surface and curated belief promotion — is specified by persona-surface.md with fourteen gates and was a seventh parallel workstream, adding an owner-edited trusted prefix row that a formed belief reaches only through explicit owner affirmation; it is complete. Milestone 23 — conversational schedule lifecycle — is specified by scheduling.md with twelve gates and was an eighth parallel workstream, adding summary-only discovery and approval-gated update, pause, resume, and terminal cancellation through the existing schedule service; it is complete. ADR-0089 separately authorizes current/recent-history native browsing and thirty-day maintenance retention for terminal schedule-owned state without adding a milestone or conversational deletion tool. Milestone 24 — SMS through the owner's iPhone — is specified by device-channel-and-sms.md with twelve gates and is a ninth parallel workstream on the same terms and is in progress, its trigger-catalog and registration-source widenings named in ADR-0081. Milestone 25 — the WhatsApp business surface — is specified by whatsapp-surface.md with twelve gates and is a tenth parallel workstream, implemented on Milestone 14's ports and in progress; its receipt-key generalization of the surface seam is named in ADR-0082.
  • Milestone 26 — Client modes and adaptive email experience: authorized on 2026-09-11 as a new independent parallel workstream under ADR-0092; email-experience.md contains the approved complete design and thirty-two gates. Twenty-seven backend and native acceptance checks are bound; four private quality gates and integrated release evidence remain pending, and the milestone remains in progress. Aggregate automatic email budgets are USD 20/day and USD 200/rolling thirty days. ADR-0096 limits automatic historical analysis to the latest ninety days for now.
  • Milestone 29 — Chat thread folders: authorized on 2026-09-16 as a new independent parallel workstream under ADR-0102; thread-folders.md declares twelve gates for flat, single-parent folders over chat conversations, owner-resolved grouping proposals from a maintenance pass, grounded model-assisted grouping with a deterministic fallback, scoped routes behind a default-off flag, and the native sidebar. ADR-0110 amended it on 2026-09-19 to admit the typed-judgment port, the TypeSafe Jev adapter behind it and an optional judgment matcher for add-to-folder proposals, all default off. All seventeen gates are bound to executable checks, and the milestone completed on 2026-09-21 when pull request 127 followed pull request 118 through hosted CI and a clean CodeRabbit review; the owner had activated folders in production on 2026-09-17.
  • Milestone 30 — Advisory approval layer: authorized on 2026-09-19 as a new independent parallel workstream under ADR-0111, the restrictive-only half of roadmap item B8; policy-and-approvals.md declares five further gates for an advisor port, a composite engine that can only escalate an allowed web or browser network read, a judgment-backed advisor that never denies, a deterministic recovery policy in the tool pipeline, and observe before enforce. All five gates are bound to executable checks and pull request 127 delivered them with a clean review; enforcement stays off in the shipped profile, and threshold calibration keeps the milestone in progress.
  • Milestone 31 — Email unsubscribe assistance: requested by the owner and shaped on 2026-09-19 as a new independent parallel workstream under ADR-0112; email-unsubscribe.md declared twenty gates for a header-derived census of bulk senders, the authenticated RFC 8058 one-click request sent through a dedicated public-HTTPS egress transport to a server-derived destination, mailto: unsubscribe, Report spam and sender cleanup as fixed label actions, owner-gesture consent for a bounded batch, scoped routes behind a default-off flag, two Chat tools, and the native Subscriptions view. It is implemented with nineteen gates bound to passing checks and reached main in pull request 131; its flag was on in production by 2026-09-21, ahead of the owner's real-mailbox smoke, and simulator UI journeys and the release evidence remain open. On 2026-10-08 ADR-0165 extended it to twenty-three gates with a suspected-spam flag that never touches Gmail and thread-level Report spam and Not spam.
  • Verified gate ceiling: Milestone 14 (289 gates).
  • Authorized workstreams: Milestones 13 through 15 in order — general-purpose subagents and delegation, inbound surfaces and pairing, operational hardening (ADR-0061) — plus Milestone 16 memory evaluation and lifecycle as an independently advancing parallel workstream whose gate ceiling cannot move ahead of Milestone 15 (ADR-0069); it completed on 2026-08-23. Milestone 17, the memory read API and the native memory browser, was authorized on 2026-08-23 as a second such workstream (ADR-0070) and completed on 2026-08-24 on the same terms, moving the verified ceiling no further. Milestone 18, first-class email integration, was authorized on 2026-08-24 as a third (ADR-0071), on the same terms again. Milestone 19, conversational schedule creation, was authorized on 2026-08-24 as a fourth (ADR-0072), and Milestone 20, calendar recurrence and conversational schedules, was authorized on 2026-08-27 as a fifth (ADR-0073), each without advancing the verified ceiling. Milestone 21, adaptive memory distillation, was authorized on 2026-08-31 as a sixth (ADR-0077) on the same terms. Milestone 22, the persona surface and curated belief promotion, was authorized on 2026-09-01 as a seventh (ADR-0079), again without advancing the verified ceiling. Milestone 23, conversational schedule lifecycle, was authorized on 2026-09-02 as an eighth (ADR-0080) on the same terms and expanded to conversational definition update on 2026-09-04 (ADR-0088). The owner authorized ADR-0089's non-milestone schedule history and retention extension on 2026-09-10. Milestone 24, SMS through the owner's iPhone, was authorized on 2026-08-26 as a ninth (ADR-0081), and Milestone 25, the WhatsApp business surface, on 2026-08-26 as a tenth (ADR-0082), on the same terms. Milestones 26 through 31 followed as independent workstreams under ADR-0092, ADR-0097, ADR-0100, ADR-0102, ADR-0111, and ADR-0112. Milestones 19, 20, 22, 23, and 29 have completed; their completion was recorded on 2026-09-28.
  • Deferred: New model-routing behavior and everything still listed in the engineering plan's roadmap subsection. Milestone 21 removes only adaptive memory distillation from that residue and Milestone 22 removes only the persona surface. Milestone 32 now admits bounded whole-store consolidation, belief merging and grounded connections; learned memory policies, semantic retrieval, external memory providers, and the remaining roadmap items stay unauthorized.
  • Project status: Milestones 0 through 13 are complete. Through Milestone 12, all 247 cumulative gates, the full local and PostgreSQL lanes, Apple package and simulator lanes, and hosted CI passed on the candidate head, and the completed integration was delivered directly to dev; code review is reserved for the final merge into main. Milestone 13 completed next: the delegation domain values and limit derivation, the ledger persistence and erasure, delegate.run and its one-transaction materializer, the child-run suspension, join, and cancel cascade, the delegation limits block with tenant admission and the default-off flag, and case 32 with the tools arm overlay are implemented, with all twenty-one gates passing locally. Pull request 64 passed all five CircleCI lanes, both Xcode Cloud archives, GitGuardian, and a finding-free final-head CodeRabbit review before merging to main, covering the implementation on that head. The capability-bound continuation passes locally. A consent-stamped failed long-research run has now been captured, re-exported under the corrected trajectory@5 projection, privacy-reviewed, and admitted as the research capability scenario. Its first live baseline attempt produced four cost_usd ceiling hits before judging at costs from USD 0.861420 through USD 0.9954325; its fifth repeat ended in a zero-usage model-stream contract failure. No repeat reached the judge, proving that the checked-in USD 0.75 per-repeat ceiling cannot produce a score on the configured balanced model. The owner subsequently approved a USD 50 evaluation budget, now encoded as USD 5 per repeat, USD 25 per suite, and USD 50 per day. The next five-repeat baseline persisted five ceiling hits at USD 7.8022225 total; the completed delegation arm also persisted five tool-call ceiling hits, zero policy failures, and USD 14.8738995 total. Those runs exposed and drove regressions for content-free model diagnostics, child-suspension polling across both queue classes, bearer-terminology false positives, and the zero-tool judge deadlock. Because delegated child usage is correctly additive, the scenario now permits 80 model and 200 tool calls while retaining every approved cost ceiling. A scored single-agent baseline then completed at scores 0.75, 0.55, 0.75, 0.65, and 0.75, with mean 0.69, floor 0.55, population variance 0.0064, zero ceiling hits, zero policy failures, and USD 7.855529 cost. It correctly remains release-blocked because one repeat is below the rubric's 0.60 floor. The provider ledger used USD 42.599510 of the initially approved USD 50 aggregate cap. The owner subsequently raised the aggregate authorization to USD 100, now encoded as the daily ceiling while the USD 5 repeat and USD 25 suite ceilings remain unchanged. A first scored delegating arm on build 877b3ec then scored 0.75, 0.75, 0.775, 0.65, and 0.35 (mean 0.655, floor 0.35) at USD 10.984242 and correctly remained blocked below the baseline. Its child ledger drove a governed default-limit and synthesis- discipline repair. The exact repaired build 3c8772a-delegating-repair scored 0.65, 0.65, 0.90, 0.65, and 0.75, with mean 0.72, floor 0.65, population variance 0.0096, zero ceiling hits, zero policy failures, and USD 11.104456 cost. Its durable suite event reports release_blocked: false, and the delegating mean exceeds the baseline's 0.69 as ADR-0063 requires. Aggregate provider spend is USD 64.688208 of the USD 100 authorization. The failed-trajectory baseline remains the comparison evidence. Pull request 72 then passed every hosted lane and a finding-free final CodeRabbit review on its final head d9827c0b, and merged on 2026-08-27 as f34d021e, completing the milestone. On 2026-09-28 agent eval gates --milestone 13 confirmed all twenty-one delegation gates and the cumulative registry, with the seven Milestone 6 sandbox gates passing in the hosted real-Docker lane. Milestone 14 is implemented and in progress with twenty-one registered gates passing locally; Milestone 15 remains authorized and specified with sixteen and has not started. Milestone 16, the parallel memory-evaluation workstream, has implemented all twenty of its gates, republished the provider evidence at formation@8, and re-recorded its baseline, and completed on 2026-08-23 when hosted CI and the CodeRabbit review loop finished clean on the dev to main pull request (merge 571f6d9); it does not move the verified gate ceiling. Its live benchmark arm ran three times — the first two failed only the absolute incomplete-runs condition, at two of 132 probe arms each, after which the harness gained a content-free failure class and one retry per probe arm, and the third published evals/capability/memory-benchmark-evidence.192a0161d881837218c0ed125c55a121663f8eda.json with four retried runs and a lift of forty-five — which is milestone evidence rather than one of the two completion conditions below. Milestone 17, the second parallel workstream, completed on 2026-08-24: all ten gates, the Python and PostgreSQL lanes, the native Apple package and simulator lanes, hosted CI, GitGuardian, and the final CodeRabbit review passed on the dev-to-main pull request (merge 3faa978), with supplemental end-to-end browser coverage later passing the same hosted gates in merge 07c8bdf. Milestone 18, the third parallel workstream, has its first-party Gmail MCP package, composition, policy, bootstrap ceremony, and monitoring recipe implemented, with all seventeen registered gates passing locally, including the four operator-managed multi-account gates added under ADR-0085. The extension reached main in pull request 94 with one CodeRabbit finding unresolved, and the owner's mailbox smokes remain outstanding. Milestone 19, the fourth parallel workstream, completed with its five registered gates once pull requests 64 and 66 passed hosted CI and a clean review. Milestone 20, the fifth, completed with its six gates on pull request 74. Milestone 22, the seventh, completed with its fourteen gates on pull request 82: the persona document, its trusted prefix row, curated promotion, the routes, the CLI, and the native editor. Milestone 23, the eighth, completed with its twelve gates on pull requests 89 and 98. Milestone 29 completed on pull requests 118 and 127. Each in-progress milestone's remaining work is itemized on the milestones page, which make docs-check reconciles against the project state.

Milestone 10A adds governed foreground skill authoring and an optional, non-joining background-review child run. Authoring stays disabled by default; tenant activation remains blocked until the evaluation threshold in the skills design passes, and that activation is roadmap item B1 rather than a Milestone 10 completion condition (ADR-0061). The provider-assisted memory extractor's evidence is accepted under ADR-0057. ADR-0068's semantic deterministic repair and retry lifecycle advanced the active policies to deterministic formation@5 and provider-assisted formation@6, and Milestone 16's admission of working-state established facts advances them again to formation@7 and formation@8 (ADR-0069). Milestone 16 republished the provider evidence at formation@8 on the intended production model and deleted the superseded formation@4 artifact, so auto again activates provider assistance for the balanced OpenAI gpt-5.6-sol and default-profile tuple and falls back safely for every other one. Tenant activation remains separate from Milestone completion under ADR-0061. The machine-readable project state records progress and evidence.

Milestone 11 was an independent, logically subsequent milestone because adding scheduling to Milestone 10 would have changed that milestone's established completion contract. Its scheduled-runs design defines a versioned schedule, immutable occurrences, deterministic civil time, bounded misfires, fresh authorization at firing, and atomic creation of an ordinary durable run; all twenty-three schedule registry entries name real checks, the production scheduler is a least-privilege role, and scheduling remains default-off until explicitly activated. Hosted CI and the final CodeRabbit review passed on head 90e9142, advancing the verified ceiling through 11.

Milestones 12 through 15 follow the pattern Milestone 11 set: a detailed-design document and an ADR land first, register the milestone's gates, and only then does implementation begin. Milestone 12 implementation is complete; its notifications-and-devices.md and ADR-0062 have landed with twenty gate.device.* and gate.notify.* entries, and Milestone 13's subagents-and-delegation.md and ADR-0063 with twenty-one gate.delegate.* entries, and Milestone 14's inbound-surfaces.md and ADR-0064 with twenty-one gate.surface.* entries, and Milestone 15's operational-hardening.md and ADR-0065 with sixteen gate.ops.* entries; no authorized milestone reports a zero row. Milestone 16's memory-evaluation-and-lifecycle.md and ADR-0069 landed the same way, with twenty further gate.memory.* entries in the existing area, and it never showed a zero row because its authorization and its specification arrived together. Milestone 17's memory-read-api-and-browser.md and ADR-0070 landed on the same day as that milestone's authorization, adding ten more gate.memory.* entries to the same area. Milestone 18's email-integration.md and ADR-0071 landed the same way on 2026-08-24, adding thirteen gate.email.* entries in a new area of their own; ADR-0085 adds four more for operator-managed multi-account Gmail. Milestone 19 reuses scheduling.md and the existing schedule area under ADR-0072, adding five gates for the deliberately narrow one-time creation bridge. Milestone 20 reuses the same design and gate area under ADR-0073, adding six gates for calendar recurrence and recurring conversational creation. Milestone 21's adaptive-memory-distillation.md and ADR-0077 add thirty-one gate.memory.* entries for integrated episodes, high-recall formation, evidence-based forgetting, comparative evidence, activation, and the attributed-communication widening of their source-grounding gate. Milestone 22's persona-surface.md and ADR-0079 landed with that milestone's authorization on 2026-09-01, adding fourteen gate.persona.* entries in a new area of their own for the trusted persona row, curated promotion, and the persona surfaces.

Milestone 12 — notifications and device identity — completed all eight build steps. The delivered slice includes the principal-scoped device registry, transactional content-free outbox, provider-partitioned claim and retry worker, APNs HTTP/2 adapter, least-privilege notify role, seven exact-scope routes, offline inbox, and native Apple registration and deep-link restoration. Review hardening made registration request identity deterministic, preserves partial per-target delivery outcomes, and executes queued navigation plus initial and changed focus behavior. All twenty gates and all 247 cumulative gates pass; PostgreSQL, Apple package, iPhone and iPad UI, unsigned Release builds, and hosted CI pass; the completed integration is delivered directly to dev. Production APNs activation was completed by the owner on 2026-08-23: a signed TestFlight build registered a physical iPhone as a production target and the authenticated test-notification route delivered the Test notification alert. That external activation followed repository completion and leaves the default-off contract unchanged for fresh deployments.

Authoritative acceptance criteria for every milestone are defined only by the canonical engineering plan; this page is a pointer, not a substitute.

Authorized work

The Milestone 10 workstreams are independently deliverable because they do not share a delivery dependency. The self-authored-skills contract is Section 30.6, the six gate.skill.* entries, and the definition of done. The memory-formation specification supplies fifteen memory-maturation gates for ordinary conversation and lifecycle, governed inspection, and the evaluation-gated provider path. The web-access tranche uses Section 32.3 and the seven formal gates in web-access.md; the browser tranche uses Section 33.3 and the ten gates in browser-automation.md. The Milestone 11 contract is its twenty-three gate.schedule.* entries plus the acceptance criteria in the engineering plan and the scheduling design.

Milestone 12's contract is its twenty gate.device.* and gate.notify.* entries plus the acceptance criteria in the engineering plan and the notifications-and-devices design; Milestone 13's is its twenty-one gate.delegate.* entries plus the plan's acceptance criteria and the delegation design, with tenant activation gated on the capability-scenario evidence; Milestone 14's is its twenty-one gate.surface.* entries plus the plan's acceptance criteria and the inbound-surfaces design; Milestone 15's is its sixteen gate.ops.* entries plus the plan's acceptance criteria and the operational-hardening design; and Milestone 16's is its twenty gate.memory.* entries plus the plan's acceptance criteria and the memory-evaluation-and-lifecycle design, whose benchmark baseline is re-recorded deliberately by every change that moves it. Milestone 17's contract is its ten gate.memory.* entries — read_api_ceiling_required, read_api_ceiling_filter, read_api_principal_isolation, read_api_pagination, read_api_filters, read_api_read_only, read_api_flag_absent, browse_contract_parity, read_api_view_projection, and read_api_error_vocabulary — plus the plan's acceptance criteria and the memory-read-api-and-browser design. Milestone 18's contract is its seventeen gate.email.* entries plus the plan's acceptance criteria and the email-integration design. Milestone 19's contract is its five new gate.schedule.* entries plus the plan's acceptance criteria and the scheduling design. Milestone 20's contract is its six new gate.schedule.* entries plus the plan's acceptance criteria and the same scheduling design. Milestone 21's contract is its thirty-one new gate.memory.* entries plus the plan's acceptance criteria and the adaptive-memory-distillation design, including ADR-0090's attributed-communication amendment to source grounding. Milestone 22's contract is its fourteen gate.persona.* entries plus the plan's acceptance criteria and the persona-surface design.

Completion rule

Milestone 10 completed after all thirty-eight Milestone 10 gates and all 204 cumulative gates passed, all required CI lanes passed on the final head, and the final CodeRabbit review had no finding or unresolved conversation. Enabling authoring for a tenant is separately governed by the rollout evidence rule and is not a completion condition (ADR-0061). Partial work does not advance the verified gate ceiling.

Milestone 11 completed after all twenty-three scheduling gates and all 227 cumulative gates passed, the PostgreSQL integration and resilience lanes passed, the hosted CI lanes passed on the final head, and the final CodeRabbit review had no finding or unresolved conversation.

Milestones 12 through 15 complete, in order, when every gate their specification declares and the cumulative registry pass, the PostgreSQL lanes pass where the milestone touches persistence, hosted CI passes on the final head, and the final CodeRabbit review is clean. The verified ceiling advances through each only after every earlier milestone has completed.

Milestone 13 completed on those terms on 2026-08-27: all twenty-one delegation gates and the cumulative registry passed, the PostgreSQL lane passed, and hosted CI and a finding-free final CodeRabbit review passed on pull request 72's final head. It was recorded on 2026-09-28, advancing the verified ceiling to Milestone 13 and 268 gates.

Milestone 14 completed on those terms on 2026-10-07: all twenty-one surface gates and the cumulative registry passed, the PostgreSQL lane passed, the owner's production Telegram smoke passed on 2026-10-06, and hosted CI and a clean final CodeRabbit review passed on pull request 153's final head, which carried the smoke's one repair. It was recorded on 2026-10-09, advancing the verified ceiling to Milestone 14 and 289 gates.

Milestone 16 completes on the same terms, with its own additional condition: the checked-in benchmark baseline equals a fresh deterministic run exactly and the provider-assisted extraction evidence has been republished at its new policy version. Being a parallel workstream changes nothing about the ceiling, which still advances only after every earlier milestone has completed.

Milestone 16 completed on those terms on 2026-08-23: every declared gate and the cumulative registry passed, the PostgreSQL lanes passed, and hosted CI and the final CodeRabbit review finished clean on the dev to main pull request.

Milestone 17 completes on the same terms, with its own additional condition: the native Apple package and simulator lanes pass, because the browser is half of what the milestone delivers and no Python gate observes Swift. Being a parallel workstream changes nothing about the ceiling, which still advances only after every earlier milestone has completed.

Milestone 17 completed on those terms on 2026-08-24: every declared gate and the cumulative registry passed, the PostgreSQL and native Apple lanes passed, and hosted CI, GitGuardian, and the final CodeRabbit review finished clean on dev-to-main pull request 58. Pull request 64 subsequently added end-to-end PostgreSQL pagination and iPhone/iPad navigation coverage with the same hosted lanes green and no unresolved review conversation.

Milestone 18 completes on the same terms, with its own additional condition: the owner's real-mailbox smoke — bootstrap consent, a scheduled triage run, an approval delivered to the phone, and one approved send — is recorded as evidence, because the mailbox is an external system no gate's fake can vouch for. Being a parallel workstream changes nothing about the ceiling, which still advances only after every earlier milestone has completed.

Milestone 19 completes when its five gates and the cumulative registry pass, all relevant local and hosted lanes pass on the final head, and the final CodeRabbit review is clean. Because it is a parallel workstream, completion does not advance the verified gate ceiling past the sequential milestones. It completed on those terms with pull requests 64 and 66 on 2026-08-25.

Milestone 20 completes when its six gates and the cumulative registry pass, all relevant local and PostgreSQL lanes pass, hosted CI passes on the final head, and the final CodeRabbit review is clean. Because it is a parallel workstream, completion does not advance the verified gate ceiling past the sequential milestones. It completed on those terms with pull request 74 on 2026-08-28.

Milestone 21 completes when its thirty-one gates and the cumulative registry pass, comparative formation@7/formation@8/formation@9 evidence for the intended production tuple is published, all relevant local, PostgreSQL, and hosted lanes pass on the final head, and the final CodeRabbit review is clean. Because it is a parallel workstream, completion does not advance the verified gate ceiling past the sequential milestones.

Milestone 22 completes when its fourteen gates and the cumulative registry pass, the deterministic benchmark baseline has been re-recorded by the change that implements snapshot de-duplication, the native Apple lanes pass because the persona editor is half of what the milestone delivers, all relevant local, PostgreSQL, and hosted lanes pass on the final head, and the final CodeRabbit review is clean. Because it is a parallel workstream, completion does not advance the verified gate ceiling past the sequential milestones. It completed on those terms with pull request 82 on 2026-09-01.

Milestone 23 completed when its twelve gates and the cumulative registry passed, the PostgreSQL lane passed, and hosted CI and a clean final CodeRabbit review passed on pull requests 89 and 98, the second merged on 2026-09-06.

Milestone 26 completes only when all thirty-two gates and approved quality requirements, relevant local/PostgreSQL/native lanes, authorized owner mailbox smoke, final-head hosted review and production delivery evidence pass. Until then its status remains in progress; it does not move the verified sequential ceiling.

Milestone 27: Bland calling and public reception

The owner approved bland-calling.md on 2026-09-11 under ADR-0097. Fourteen gates are registered. This independent workstream remains in progress. Calling has been active in production since 2026-09-17, and every pull request that carried it to main passed hosted CI and a clean review.

Local code now implements the scoped tool rosters, approval-bound dispatch, public correspondence intake, retained owner results, erasure and call-result notifications. Twelve gates have executable offline bindings; provider admission and live release remain pending, and binding them will set the final head that the exact-head review needs. Setup covers the remaining owner configuration and external verification.

Milestone 28: People and relationship memory

The owner authorized implementation on 2026-09-15 under ADR-0100. People and relationship memory specifies 36 gates, stable identities, temporal relationships, source-linked history, governed formation/retrieval, correction/erasure and the native People surface. This independent workstream is in progress and does not move the verified ceiling; none of its thirty-six gates is yet bound to an executable check.

ADR-0101 removes evaluation-dependent People availability; the complete implemented feature is enabled by default while quality measurement continues.

Milestone 29: Chat thread folders

The owner authorized implementation on 2026-09-16 under ADR-0102. Chat thread folders specifies twelve gates: flat, single-parent folders as server-owned session state; owner-resolved grouping proposals with a bounded open set, durable decline and withdrawal; grounded model-assisted grouping with a deterministic lexical fallback; nine routes on the session scopes behind a default-off flag; and native sidebar folder sections with proposal review. The implementation landed on 2026-09-16 with every gate bound to an executable check, and pull request 118 carried it to main with hosted CI and a clean review; the owner activated it in production on 2026-09-17.

ADR-0110 amended the milestone on 2026-09-19. Typed judgment specifies a provider-neutral port for closed, typed questions with the TypeSafe Jev adapter as its first provider and declares four gate.judgment.* gates; Chat thread folders gains a thirteenth gate for an optional judgment matcher that decides which existing folder an unfiled conversation belongs in. Both are default off, the matcher only ever proposes, and any failure returns the existing grouping unchanged. The same ADR admits an offline, non-activating email-importance evaluation under Milestone 26. The port, the adapter, the fake and the selector landed on 2026-09-20 with the four gate.judgment.* gates bound to executable checks, and the judgment matcher landed the same day with gate.folder.judgment_matching bound to a composed end-to-end check. No judgment request is made in a deployment until the owner sets the selector, the key, and the matcher's knob. Pull request 127 carried the amendment to main with hosted CI and a clean review on 2026-09-21, which completed this independent workstream.

Milestone 30: Advisory approval layer

The owner authorized implementation on 2026-09-19 under ADR-0111, the policy ADR roadmap item B8 requires for its second half. Policy and approvals specifies five further gates: the composite engine cannot lower a rank; the advisor runs on allow paths only and once per invocation; an unavailable advisor abstains; the advisor is blind to the rules and sees only redacted, delimited arguments; and the layer off or observing changes no decision. The first advisor uses the typed-judgment port rather than the model gateway, a divergence ADR-0111 records, and the tool pipeline gains a deterministic recovery policy. General standing approval grants stay on the roadmap. The implementation landed on 2026-09-20 with every gate bound to an executable check. Observing is the environment flag AGENT_POLICY_ADVISORY_OBSERVE_ENABLED, which moves no policy version; enforcing is the profile value advisory.enabled, which does and therefore unbinds the memory-formation release evidence until it is regenerated (ADR-0111, amended 2026-09-20). Pull request 127 delivered it with hosted CI and a clean review; threshold calibration and the final-head review remain. This independent workstream is in progress and does not move the verified ceiling.

Milestone 31: Email unsubscribe assistance

The owner requested this capability on 2026-09-19 and decided its surfaces, consent model, egress and fallbacks under ADR-0112. Email unsubscribe assistance specified twenty gates: a deterministic census of bulk senders read from header metadata inside the existing foreground refresh; the RFC 8058 one-click request, offered only for mail Gmail authenticated and sent by one builtin tool whose destination is server-derived and never an argument; a dedicated public-HTTPS egress transport for that tool alone; the header's mailto: message through the account's send server; Report spam, Not spam and sender cleanup as fixed label deltas; owner-gesture consent for a batch of at most twenty-five senders and one by-value approval per batch in Chat; four routes behind a default-off flag; and the native Subscriptions view and thread action. The owner directed implementation on 2026-09-19. Nineteen gates bind passing checks, and pull request 131 delivered it with hosted CI and a clean review; its flag was on in production by 2026-09-21. Simulator UI journeys, the owner's real-mailbox smoke and integrated release evidence remain open. On 2026-10-08 the owner extended the milestone under ADR-0165 with three more gates: a spam verdict in the importance assessment, informed by Gmail's own DMARC result, flags suspected spam without touching Gmail; the owner clears a flag or reports one thread as spam with Archive's gesture, through two routes under /v1/email/threads behind the same flag. This independent workstream is in progress and does not move the verified ceiling.