v0.10.0

This milestone completes the clean-cut transition to an agent-first team wiki: every capture is immutable evidence, one bounded librarian files what it establishes, code enforces the contract, and every reader sees only knowledge within their audience.

⚠ BREAKING CHANGES

• Replaced the previous capture, queue, knowledge, and operational contracts with the single team wiki runtime contract. Deployments carrying pre-milestone test data must reset that data before adopting this release.
• Moved audience selection to the capture door, removed ops/acl.json, and standardized principals on lists of groups. Existing deployment identity and channel files must use the new grammar.
• Removed the superseded review, repair, generated-view, meeting, Drive, entity, cron-control, and compatibility surfaces.

Added

• Exact binary evidence, typed extraction, selective OCR, immutable readable source pages, and local private acquisition through the MCP bridge.
• First-class cited contradictions, explicit resolution provenance, exact diffs, autonomous deterministic maintenance, and reader-scoped entity descriptions.
• Exact cross-capture identity convergence, deployment liveness, and explicit index convergence health.

Changed

• Rebuilt Stigmergy around one kind-free capture envelope, one durable queue, one serialized Git writer, and one bounded librarian.
• Replaced entity dossiers and generated views with opaque identities, scoped sourced claims, and deterministic registry projection.
• Standardized runtime model access on the closed OpenRouter allowlist.

Fixed

• Made contradiction resolution work through direct and bridged capture while requiring a coherent page update.
• Reconciled librarian reports with landed diffs and surfaced skipped operations without leaking restricted material.
• Converged identities across scopes and anchored pages to every unambiguous exact entity mention.
• Preserved complete Slack chronology and readable evidence while storing identical attachment bytes once.
• Bounded model, embedding, database, worker, index, and deployment failure paths.

Removed

• Legacy capture kinds, generated knowledge views, approval/task states, standalone repair workflows, and user-like operational CLIs.
• Compatibility and migration machinery for the abandoned test-only contract.

v0.9.1

Fixed

• The repairs status migration aborted on any database carrying a retired row, and the app
  crash-looped. ALTER TABLE ... RENAME carries a table's CONSTRAINTS with it, so a database
  upgraded to 0.9.0 still had repair_proposals_status_check — which permits
  pending|approved|rejected|applied|failed and does NOT permit skipped. The migration's own
  UPDATE ... SET status = 'skipped' violated the constraint that was still standing, before the
  swap that replaces it ever ran; stigmergy-server exited 2 on every start with "cannot read the
  index (CheckViolation)". The legacy constraint is now dropped BEFORE the migration writes a value
  it does not permit — three steps, not two, because an UPDATE has to happen in between
• the test covering that migration dropped the constraint first and its docstring claimed "that is
  the state the old release left the column in". That sentence was false, and it is why 0.9.0
  shipped green: the test constructed the one state in which the migration works. Its twin now
  starts where a real upgrade starts

v0.8.0

An entity is born written, and keeps being written (ADR 042, #131).
Twelve of the first brain's nineteen entity pages said nothing about the entity: the two hand
doors rendered the template with the name filled in. There is no deterministic birth any more — a
steward's registration is a capture, the librarian writes the page from what the steward said and
what the brain already holds, and a filing that establishes something about a known entity grows
its page.

⚠ BREAKING CHANGES

• stigmergy-entities create commissions a capture instead of committing a page: it needs
  --about (what the entity is, in your own words), the database (--dsn) and the evidence
  environment the drop CLIs use, and prints the capture to follow; the page appears when the
  capture files, born confirmed by --by (or your git email)
• POST /admin/api/entities/create takes about (required) and no role, and answers the queued
  row (id, status: queued, entity_id, name, message) instead of a commit; the console's
  Register form asks "What is it?" and opens the capture
• server.review.create_and_record is gone (commission_registration replaces it);
  entities.mint and entities.remote.mint_via_clone are gone (entities.guard keeps the two
  shared refusals; decide_via_clone is the one server-driven door)
• brain_submit refuses the four register_* hints from every client door
• the knowledge repo's briefs change with it (e118c8a): look before you write, the registration
  paragraph, entity_updates

Added

• entity_updates in the librarian's account (both flows): what the material establishes about a
  registered entity is APPENDED under that page's own ## Facts / ## Connections, updated:
  moved, lines the page already carries skipped, the file proved byte for byte; the report says
  "It adds N facts and M connections to the page of id"; refusals update-unknown-entity and
  update-of-new-entity
• a steward's registration through the librarian: capture.schema.registration_hints /
  registration_from_hints, the brief's REGISTRATION paragraph, identity/registration-missing
  for an account that ignores it, the identity gate's not-confirmed-by-its-steward, the ledger
  row written by the worker after the push
• birth.render_page refuses an entity page with no What / Who, drops a section with nothing to
  say, and strips the template's HTML comments — nineteen pages had carried them, indexed as text

Removed

• the deterministic mint: entities/mint.py, remote.mint_via_clone, review.create_and_record,
  the console's synchronous Register and the CLI's create commit

v0.7.0

File first, govern after (ADR 041, #125).
A capture never waits on a person any more. A name the registry does not know used to park the
capture on a question to its submitter and then on a steward; two of five real notes were cancelled
by their own authors that way. Now the librarian files at once and PROPOSES the entity — a complete
page with approved_by: "" and a proposed registry entry — and a steward approves, merges or
declines it later, from whichever door is nearest, in one governed commit.

⚠ BREAKING CHANGES

• the capture statuses are queued · claimed · filed · rejected · failed (resolved survives
  read-only on old rows). needs_input and triage are retired words the queue refuses by name;
  rows found in them are returned to queued once at startup, so nothing already captured is lost
• the brain_reply MCP tool is gone — nine tools, pinned by a test. brain_submit's
  acknowledgement now names the entities the capture will be filed against and says unknown ones
  will be proposed
• review_queue / review_decide speak two new item kinds, identity-proposal (item id = the
  entity id; verdicts approve, merge with into, decline) and alias-proposal (item id =
  <entity id>:<alias>; approve, decline); entity-proposal and parked-capture are read-only
  legacy kinds on the rows that carry them
• stigmergy-queue keeps list · show · claim · reclaim · purge and loses requeue · resolve · reject; stigmergy-entities is now pending · approve · decline · merge · create · regenerate
  (propose is the librarian's job)
• the admin console's queue/{id}/requeue|resolve|reject routes are gone; entities/{id} takes a
  registry id; entities/decide and entities/create are new. The Captures page is read-only
• entity pages carry approved_by (absent = confirmed before the field existed, "" = proposed,
  a name = who confirmed it) and proposed_aliases; the registry carries proposed,
  approved_by and proposed_aliases. The knowledge repo's librarian and meeting-distiller briefs,
  entity template and linter move with it — both repos ship together

Added

• librarian.identity — the proposal writer: folds every new name against the registry with the
  birth gate's own fold (a known name becomes a proposed spelling, never a twin), writes the
  entity page from every field the reasoning filled, reads the ledger so a declined name is never
  proposed twice, and tells the gates exactly what it wrote
• the ninth gate, identity: every write under wiki/entities/ is a declared proposal arriving
  unconfirmed, or a proposed spelling proved byte for byte — nothing else
• entities.decide — approve / merge / decline for an identity, approve / decline for a spelling,
  one apply with preflight, drift refusal, secrets scan, one commit and rollback on a failed push;
  remote.decide_via_clone for the deployed doors, Decided-by: trailer
• the inbox is derived from the registry, pages_index and the ledger — no new table; proposed
  entities are visible in search and list_entities, marked
• the console's Entities desk: each proposal with the registry verdict on its name, merge
  candidates, Approve / Merge into… / Decline, proposed spellings, the registry browser, and
  Register an entity born confirmed with the live name check
• the Slack card for a proposal, with the same three verbs
• the filing report's proposals clause; the digest counts proposals decided

Removed

• the ask-back: brain_reply, the needs_input/triage states, the three dispositions, the
  parked-capture mint door, entities.situations, the console's queue drain, the meeting flow's
  "reuse the parked distillation" rule (nothing is re-filed, so nothing is re-read)

v0.6.0

The admin console grows up from a quick ops skin into a control room a steward or an admin can
read without the runbook open: grouped navigation, a unified inbox, plain-language labels for every
system word (the raw word one hover away), a "how to read this page" explainer per page, and
charts — every one with a table twin — in the README's own "colour is who decides" key, validated
for colour-vision deficiency in both modes — which a steward can now choose between, or leave to
the device.

Added

• /admin/api/inbox — everything parking on a human as ONE list, the same read the Slack
  doorbell rings from (server.review.items_for_doorbell), with per-kind counts; the sidebar badge
  is its count
• /admin/api/entities/registry and /admin/api/entities/resolve — the registry this server
  serves (the index's snapshot, else the --entity-registry file) and the pre-mint check over it
  with the mint gate's own folds: registered (requeue, nothing to mint), collides (the gate
  will refuse it — alias it in the knowledge repo instead of minting a twin), similar (advisory),
  clear. Both entity routes attach a verdict per unresolved name; the Approve form checks the
  Name and every Alias live as the steward types; the Entities page carries a searchable registry
  browser
• /admin/api/metrics?days= — captures by arrival day and outcome (queue.outcomes_by_day, new
  beside counts_by_status), capture→filed samples, ask outcomes per day shaped with the pilot
  report's own predicates, calls per day/tool/identity, each job's run history, the latest
  decisions, repair counts
• /admin/api/meta ships every closed vocabulary the console renders (statuses with their parked
  and terminal subsets, situations, repair kinds, severities, item kinds, decision doors), so the
  frontend never hardcodes a list that could drift
• the Dashboard's live write path — the window's captures flowing through the model's draft and
  code's gates into landed / parked / refused / could-not-finish, with real counts
• an appearance picker — Auto, Light or Dark — in the sidebar and on the login screen, remembered
  in the browser and stamped before the first paint (no flash on the way in). Every colour token
  is one light-dark(light, dark) declaration, so the two themes cannot drift apart

Changed

• the console's pages: Dashboard, Inbox, Captures, Entities, Repairs, Gardener, Index, Worker,
  Jobs, Digest, Activity (the old tab names overview/queue/crons still route); every page
  opens with a collapsible explainer, pages with a time axis share one 7/30/90-day window
• the frontend is one module per page under static/assets/views/, with copy.js (the
  vocabulary), charts.js (SVG charts built with createElementNS) and state.js; the static
  discipline tests follow the split
• the Jobs page renders no levers at all without the GitHub token, instead of disabled ones

Fixed

• the shell and its assets carry cache-control: no-cache — a deploy that renames a module no
  longer leaves a browser running the old app.js against new imports for hours (a blank page)
• inline style attributes, which style-src 'self' silently refused, are gone: every style goes
  through the CSSOM
• the librarian's placeholder for a park that named nothing (something unnamed) is refused by
  value at the terminal name gate (entities.birth) — every mint door, not only the prefill rule;
  entities.situations.is_mintable_name is the one comparison, and the console's per-name checks
  carry it as mintable so no surface offers a button for it
• metrics runs off the event loop and every read of a table that only grows is bounded in SQL:
  decisions.recent_decisions (the ledger feed), repair.store.counts_by_status (the whole-table
  histogram), a ceiling on the pending proposals the console reads, pilot_report.answer_shape_by_day
  (the report's own classifier, grouped in SQL and pinned against the Python original)
• the pre-mint similarity listing folds the registry once per request, not once per name
• answer_shape_by_day's SQL mirror of shape_of no longer casts result ->> 'citations' to
  int. audit_log.result is JSONB with nothing under it, and an ask row older than
  audit_summary carries citations as the LIST of page paths — so the cast raised on real data
  while every test that fed it today's integer stayed green, and since eight of the console's
  eleven pages fetch metrics, one legacy row rendered almost the whole console as failed. The
  mirror now asks the question shape_of asks — truthiness — as a JSONB comparison against the
  falsy set, which cannot raise whatever a past writer left in the column
• the deterministic-check count is gone from the prose that nothing pins: six sites said "nine"
  and the live workflow said "eight" for a tuple of ten. The package docstrings, the console's two
  copies of the cron description and the cron template now say "the deterministic checks", which
  survives the next one being added; the sites that state the number derive it. Beside it, a guard
  that every *.py a code map names in backticks exists

v0.5.0

v0.4.0's follow-up tracker, emptied: the five issues the open-models port filed against itself (#110–#113, #115), each landed behind the instrument that owns it. The reference deployment's embeddings move to qwen3-embedding-8b at 2560 MRL dimensions through the same OpenRouter key — gated by the retrieval golden holding recall@5 at 0.969, identical to the text-embedding-3-large baseline it replaces.

Added

• EMBED_DIMENSIONS — MRL truncation for over-ceiling embedding models: request-level dimensions, sent only when set, refused by name when malformed; what fits a 4096-native model under the schema's 4000-dimension HNSW ceiling (#116, closes #115)
• index_meta records the embedding HOST beside model and dim, and search_arms refuses a host mismatch by name BEFORE the first embedding — the same model name on two hosts is not provably the same vector space. Legacy indexes without a recorded host skip the check until their next rebuild (#117, closes #112)

Fixed

• the vision OCR pass reaches cost_usd: both forms return token usage as data, the pass is priced through the librarian's one table (an unpriced vision model degrades to a loud $0.00 line, never a refused capture), and the report names the share on conversion_cost_usd (#118, closes #110)
• the visibility-lease derivation carries the drive conversion budget, imported from the kernel's own vision clocks: 1290s at the class default, 1890s on the deployed worker, every prose statement moved in the same commit, the reclaim refusal's example pinned by a parity test (#119, closes #113)
• the docs-claims environment guards see the model-seam env family through one name-enumerated pattern (#120, closes #111)

v0.4.0

The open-models port. Every model seam now runs on any pydantic-ai provider — OpenRouter first — and the reference deployment moved to open-weight models behind this repo's own instruments: the filing golden passes every bar on deepseek-v4-flash at roughly $0.008 per capture, the QA golden scores 1.00 on every axis on glm-5.2, and the retrieval golden holds recall@5 at 0.969 with the same embedding model served through the new host. No gate, ACL, fence or eval moved: the seams did.

Added

• ask takes the two-form model convention — a bare ANSWER_MODEL stays the OpenAI Responses API; a provider-prefixed pydantic-ai id authenticates with that provider's own key — and openrouter joins the librarian's key preflight and pricing table (#107)
• the embedder speaks to any OpenAI-compatible /embeddings host (EMBED_BASE_URL, EMBED_API_KEY, EMBED_MODEL), with the recorded-model rule keeping every query in its index's vector space; the worker boot strips the embed credential like the OpenAI one (#108)
• two-form vision OCR: a provider-prefixed VISION_MODEL transcribes poppler-rasterized page images through pydantic-ai — bounded pages, a spoken cut — while the bare Gemini form stays byte-for-byte what it was (#109)

Fixed

• an empty model-chosen search query is a repairable refusal handed back to the asking agent, instead of a provider 400 that crashed the whole ask (#114)
• the audit sweep over the port (#114): the OpenAI key never travels to a non-default embedding host; a prefixed VISION_MODEL missing its provider's key refuses naming that key instead of advising a requeue that could never work; the rasterizer carries timeouts and a pixel bound against raster-bomb pages; the provider→key table lives once, in the kernel; the worker boot says out loud when one credential wears two names. Deferred halves filed as #110–#113
• account schemas decode nested structures a provider's tool-calling stringifies — the defect that filed meetings with decisions empty on routes returning nested lists as JSON strings (#114)

v0.3.1

The three decisions v0.3.0's adversarial review filed (#101, #102, #103), each taken on the side
of simplicity: no new pass machinery, one new knob where two siblings already had theirs, and two
of the three fixes ride seams that already existed.

Fixed

• The editorial sweep is bounded on every axis — and deliberately not by batching (#101). Its
  checks are about PAIRS, and a batch boundary would silently decide which contradictions are ever
  visible. Instead: the changed half keeps the newest STIGMERGY_GARDENER_SWEEP_CHANGED_CEILING
  (default 30) filings, every fenced body is clamped, and the prompt becomes settings-shaped —
  never corpus-shaped, even on a first run or after a cron outage. The overflow is counted, named,
  and never lost: it joins the unchanged pool the rotating sample already covers. A failed night
  re-presents a bounded population, so the frozen-watermark loop degrades to a bounded retry.
• The view sweep yields to arriving work, and a hung synthesis has a clock (#102).
  should_stop becomes a reason-string contract: the worker's own callable says "the process is
  shutting down" or "a capture is waiting in the queue", and the recorded deferral repeats those
  words — a capture submitted mid-pass now costs one entity's regeneration, not a whole ceiling's.
  SYNTHESIS_TIMEOUT_S turns a provider that stops answering into the existing withheld shape
  instead of a hung worker loop with no lease, no row and no log line.
• The night's one number bounds the asks, not only the inbox (#103). The per-finding repair
  roads (body, merge) stop at max_proposals_per_run asks with a recorded ask-ceiling-reached
  reason — a night of declined drafts, which store nothing by design, is now a bounded bill. The
  recurrence stays deliberate; what is bounded is its nightly cost, and the spend records from
  #81 show how close a night comes.

v0.3.0

The governed repair loop grows from a design into a working subsystem with four proposal kinds,
the view layer converges itself, and the ops/ control files a deployed process trusts stop being
deploy-time copies. Two adversarial audits ran over the batch — one on the original five issues,
one on the whole window — and their surviving findings are either fixed here or filed with a
decision recorded (#101, #102, #103).

Added

• The governed repair loop (ADR 039, #69/#71/#72/#89): a gardener finding gets a path to zero
  through four proposal kinds — additive edits, one drafted entity-body per page, governed
  delete (the one kind no model may propose), and entity-alias (two registry entries that are
  one entity: the model picks the survivor, code computes the sweep, one steward decision per pair
  permanently). A MODEL proposes, CODE validates twice, a HUMAN approves one at a time.
• The view sweep (#86): a view is never stale, whatever wrote the corpus — a state-based
  convergence pass on the librarian worker's idle branch, one commit per entity, cooperative
  shutdown between entities, mutual exclusion by advisory lock.
• Two more model checks (#84, #89): an entity body that is written and says nothing, and two
  registry entries that denote one company; plus a tenth deterministic check
  (anchored-to-superseded-entity, #88) counting the residual an applied merge cannot sweep up.
• The ops-file cache (ADR 040, #74/#79): the entity registry, the identity roster and the
  channel scope map ride the derived index as verbatim-TEXT snapshots, refreshed by the push
  webhook (fetched at the BRANCH ref — a replayed delivery can only install what the branch says
  now) and reconciled per file by the nightly rebuild (the access files are never cleared over an
  absent checkout copy). A revocation lands within seconds of its push instead of at the next
  deploy. The page road gets delivery-id replay protection; an EMPTY snapshot resolves nobody.
• The budgets' feedback loop (#75/#81): every proposer model call records requests and tool
  calls against its limits — and token counts — into job_runs.stats; the edits budget derives
  from the batch it carries; kernel.llm.model_override is the public seam for proving tool-loop
  properties against the real agent, keyless.
• Agentic entity resolution (#77): a filing near miss is the agent's judgment, with four new
  filing-eval fixtures pinning it.
• The meeting distiller sees the corpus it files into (ADR 038); project joins the entity types
  (ADR 037); every review decision records its door, and stale Slack cards close themselves.

Fixed

• An empty entity-body draft is the park, not a validation error — one model call, not two,
  forever (#83).
• Every unfenced prompt-header scalar is hygiened, not only the path — and the seventh prompt
  builder (views/synthesis.py) that the consolidation sweep never reached (#92 follow-up).
• A backlink that stopped qualifying stops being cited (#85); the view agent names its model
  instead of inheriting the read path's (#90).
• The non-additive repair kinds never rebase: a lost push race fails clean instead of landing a
  diff the gates never judged against that base (#88).
• The weekly digest reads every model pass's error, so a partial run can never render as a
  clean one; a blank entity body no longer falls between the deterministic and the model halves
  of the empty-body pair.
• STIGMERGY_REPAIR_BATCH — the one count knob that multiplies a model budget — gains a maximum;
  entity names refuse control characters; three entities refusals stop echoing a foreign
  exception; non-ASCII stems are written verbatim in frontmatter lists.
• The registry served by MCP follows main, not the deploy (#74); stigmergy-index --check
  lints the copy the server actually serves.

Changed

• The whole-branch refactor sweep (#92): one definition per fact across fourteen modules —
  shared seams single-sourced, dead surface dropped, the ordinary/meeting filing copy-paste
  collapsed, prompt-scalar hygiene folded down into stigmergy.text.
• _expansion_terms is bounded by count and term length; the two registry parsers' authority is
  stated (strict for writing, tolerant for serving); the admin console's Index panel answers
  freshness for all three cached ops files.

v0.2.2

Prose only. No executable code changed in this release, and that is a mechanically verified claim rather than an assurance: for all 156 modified modules the AST with docstrings blanked is byte-identical to v0.2.1, so no statement, import, name or string literal moved.

Changed

• Narrative prose purged from code comments and reference documentation. This repository's own doctrine — a comment states a constraint the code cannot show, never the story of how the code got here — was not being honoured by the code itself: essay-length docstrings, obituaries for deleted subsystems and process narration had accumulated to roughly 46% of src/. That text cost twice: every change paid upkeep on prose no test protects, and a reader (agent or human) spent most of a module's context on history instead of on the system.

  	before	after
  src/	41,180 lines	29,946
  src/ comment + docstring lines	~19,000	7,824 (−59%)
  docs/reference/	6,687 lines	5,798
  evals/README.md	495 lines	229

  Every package's index.md code map is rewritten as a present-tense map. Nothing is lost that was not already recorded: history lives in git log and in docs/decisions/, which is exactly what makes it safe to remove from the code.

• Every trace of subsystems that no longer exist is gone from code prose — the learning loop, the extraction pipeline, the retired filing backend, the read site and the canon lane. A reader grepping src/ for a name now finds only names the system actually has.

Notes for contributors

Four bodies of prose were deliberately left alone, each for a reason worth knowing before the next cleanup: tests/ (a test's prose is this project's specification, and the reproduce-first rule requires a test to say what the old behaviour was), docs/decisions/ (an ADR records a decision, so history belongs there by design), MCP tool and pydantic schema docstrings (they are contracts sent to models and clients, not commentary), and the frozen fixtures asserted byte-identical to the knowledge repo's own copies.

Two mechanical guards caught genuine over-deletion during the work — the meeting brief's date-bearing-body-link contract marker and the operator runbook's sweep line — and both were satisfied by restoring the fact, never by relaxing the check.

Verification

4,565 tests pass at 94.24% coverage against real Postgres, real MinIO and real git; make lint clean over src tests evals scripts.