v0.99.0

0.99.0 - 2026-07-10

Added

• nika:image_fx — deterministic artistic effects, the 26th builtin
  (stdlib §Media graduate #3). The image editing deferred row comes
  home zero-dep: 15 op families (dither · palette · duotone · pixelate ·
  halftone · grain · vignette · chromatic aberration · scanlines ·
  glitch · ascii …) over a hand-rolled PNG codec (full RFC 1951 dynamic
  Huffman inflate · 5-filter decode · CRC), seeded and byte-identical —
  the recipe rides the PNG tEXt chunk (image_fx/v1), the artifact
  sha256 rides the trace chain, and a re-run with the same inputs
  idempotently skips.
• nika:chart — deterministic chart artifacts, the 27th builtin
  (stdlib §Media graduate #4). Rows + a semantic spec compile to
  byte-identical SVG (sha256 → trace chain): five closed types (bar ·
  line · area_band · scatter · heatmap), typed semantics (usd ·
  duration_ms · timestamp · category …), out: must end in .svg (the
  attestation surface) and compile_to: vega_lite writes the .vl.json
  sibling. Zero dependencies; parity proven byte-exact across
  architectures (wasm32-wasip1 ≡ aarch64).
• nika check --model <provider/model> — the static preview of the
  run override: the envelope is re-priced AS IF the flag replaced the
  file's default (per-task model: still wins), so what check shows IS
  what run will refuse or allow.
• Egress to outputs — the workflow boundary earns its valve. The
  capture-taint law is deliberate (the provider saw the key — its
  response is not provably clean), but a workflow that calls an
  authenticated API and RETURNS the result had no sanctioned path: the
  embedded api-upload-and-create template failed its own audit (the
  night battery's catch), documented as the one known gap. The gap
  closes the way its own note asked: egress: [{ to: "outputs" }] on
  the secret declassifies the workflow boundary itself — sink-only,
  secret-specific, never authorizes a send, default-deny when absent
  (spec 01-envelope §egress). KNOWN_GAP is empty: every embedded
  template now passes its own audit, with zero exceptions.
• The MODELS rung — every model: must resolve in THIS binary. The
  ladder validated tools but never models: a vendor-cataloged provider
  the resolver cannot drive (azure/…) and a bare model id
  (gpt-5-turbo) both audited green — the bare one even wore a
  conjured price. Both are findings now (exit 2) with the fix taught
  in-line; pricing refuses to price what cannot resolve (unpriced beats
  conjured); the --json payload carries models_resolve +
  model_findings[]; and the SAME law guards the MCP nika_check lane
  (nika_providers::resolve_refusal — the two machine lanes cannot
  disagree).
• nika wire learns opencode and hermes — the two ecosystems
  that natively read what nika init writes get first-class MCP wiring
  (wire all now covers 8 targets). OpenCode: project-local
  opencode.json, its own mcp.nika shape, idempotent merge. Hermes:
  ~/.hermes/config.yaml under the Zed contract — create when missing,
  recognize current, otherwise hand back the exact snippet and leave a
  foreign YAML byte-identical.
• OpenRouter calls carry the app-attribution pair (HTTP-Referer: https://nika.sh + X-Title: Nika) — runs surface as Nika on the
  public rankings instead of an anonymous key. Openrouter-profile only;
  peers may 400 on surprise headers (proven both directions).
• The plugin ships three slash commands (/nika:check ·
  /nika:explain · /nika:new) — born under .agents/plugins/nika/
  per the marketplace mirror law; the commands read the --json
  payload, not the prose.
• nika explain --forecast — learned truth before a run. Duration,
  cost and risk priors computed from YOUR local traces (.nika/traces/)
  — deterministic stats, never a model call, never the network. The
  honesty ladder is a type: never-run says so · one run is « last run » ·
  2-4 runs earn a min–max range · p50/p90 bands are earned at n ≥ 5
  (Hyndman & Fan type-7, the numpy/R default). Costs compose the ≥
  floor whenever unpriced spend participates — absence stays a dash,
  never $0. Retried-then-passed counts as flaky, never failed; cache
  hits and other-model runs are excluded from bands and named. The
  section auto-appears in nika explain <file> once 3 runs exist; the
  flag forces it, and --json gains a versioned forecast key
  (internally-tagged rungs — consumers tolerate unknown kinds).
• The missing-input trap is taught at check time — nika check
  prints an [inputs] HINT when a nika:read path that resolves
  statically (a literal, or one ${{ vars.X }} with a literal default)
  does not exist here; never an error — the file may appear at run
  time, and anything dynamic is never guessed
  (nika_schema::check::static_read_paths, the pure half).
• The release tarball is funnel-gated — scripts/ci/funnel-e2e.sh
  plays the stranger's first path against the EXACT tarball about to
  ship (clean HOME · offline · content asserts); a broken first-run
  never uploads.

Fixed

• The run card advertises the full 64-hex chain head — trace verify printed the whole sha256 while the run card truncated to 32,
  so the taught receipts loop could only prefix-match. Byte-exact ==
  now closes it (CI-assertable).
• The broken editor modeline names itself — cause, not symptom. A
  weak copier de-comments the # yaml-language-server: line; YAML then
  fails at the first mapping (« line 14, nika: v1 ») while the fault
  is line 1 — repair loops chased the symptom forever (0/13 measured on
  a 14B grid). Both forms now teach the fix on the offending line, and
  the class is mirrored spec-side as conformance fixtures 014/015 —
  writing them un-crashed the oracle's scan-failure path.
• The cost floor prices the EFFECTIVE model. The delegation idiom
  agents are taught (--model <p/m> --max-cost-usd <usd>) never met
  the pre-start refusal — the floor was computed from the file's model
  while the run used the override. The budget preflight now re-prices
  the effective envelope; the mock-override preview idiom still passes.
• The PNG heatmap speaks the SVG's quantized bins — the design pass
  quantized the SVG cells onto the legend's 8 shared bins while the PNG
  projection of the same recipe kept the continuous ramp; one shared
  fill law now feeds both surfaces (every legend swatch IS a color a
  cell can wear).
• Fold verdicts are terminal kinds only — a journal truncated after
  its opening line no longer surfaces workflow_started as if the
  crashed run had reached a state; nika context reads it as
  honestly-unknown.

Changed

• The chart design pass — per-mode palettes, computable, never
  eyeballed. The six-check validator refuted « Okabe-Ito is CVD-safe
  on both modes » (4 slots outside the dark lightness band · the yellow
  at 1.29:1 on white): dark becomes a SELECTED palette (same seven
  hues, its own steps, all checks green against #0f1318; light
  re-steps one slot, yellow → gold), series and diverging bins ride CSS
  classes through the prefers-color-scheme seam (ONE byte-stable
  document renders both themes), heatmap cells quantize onto the
  legend's 8 shared bins with a 2px surface gap, bars cap at 24px with
  a rounded data-end and square baseline, and the diverging midpoint
  goes near-surface graphite in dark — « no change » must recede, never
  glow. A BOTH-lists-style test guards the style block against palette
  drift.
• nika-dap::stats::conformal_upper — the forecast's first THEOREM.
  A distribution-free, finite-sample upper prediction bound (split
  conformal, order-statistic form): for exchangeable runs the NEXT one
  falls at or below the k-th order statistic with probability ≥ k/(n+1),
  exactly k/(n+1) for continuous data (arXiv:2411.11824 Theorem 3.2). The
  level arrives as a rational (the f64 route mis-computes the exact
  feasibility frontier — 0.9·10 ceils to 10); nine runs earn a
  guaranteed 90% bound, nineteen earn 95%. Proven by a deterministic
  leave-one-out property test that COUNTS the theorem, never samples
  it. Renderer wiring lands with forecast-R3.
• nika-dap is its own crate — the trace-forensics plane has one
  home. The DAP replay server moves out of nika-cli (the crate sat
  at 98.9% of its size cap before the forecast landed), and the seams
  every forensic reader shares descend with it: the tolerant NDJSON
  reader (recover_events), the tamper-evidence chain walk
  (chain::walk + ONE CHAIN_GENESIS the sink now imports — three
  private sha256 copies unified), and the source-identity hashes.
  nika-cli re-exports every seam at its old path — zero behavior
  change, nika dap answers exactly as before.

---

What's Changed

• docs(changelog): append v0.98.0 by @github-actions[bot] in #308
• feat(nika-cli): the DAG draws itself in the terminal — real wires, never a lie by @ThibautMelen in #309
• feat(nika-cli): wire zed — context_servers with a JSONC-honest fallback by @ThibautMelen in #310
• feat(nika-cli): trace retention — bounded by default, never silently (ADR-100) by @ThibautMelen in #311
• fix(nika-cli): doctor sees zed — the wire target it could not diagnose by @ThibautMelen in #314
• feat(nika-event): task_recovered — the repair becomes readable in the kind stream by @ThibautMelen in #313
• feat(nika-cli): doctor + explain speak the display seam by @ThibautMelen in #315
• feat(nika-cli): nika context — the whole workspace truth in one call by @ThibautMelen in #312
• docs/readme mirror verbs by @ThibautMelen in #316
• docs: readme freshness, manifesto emergence, first-run feedback forms by @ThibautMelen in #317
• docs(readme): brand header — the lockup hotlinks the canonical kit by @ThibautMelen in #318
• ci(coherence): pack CONTENT pin + agents-mirror pin join the nightly by @ThibautMelen in #322
• docs(readme): the constellation — every sibling repo, one table by @ThibautMelen in #329
• fix(hygiene): vector 9 follows the org storefront — canonical tap name + the two new repos by @ThibautMelen in #336
• refactor(nika-catalog): session-2b splits along its rule families by @ThibautMelen in #326
• fix: the V-arc findings — the stranger's first path holds by @ThibautMelen in #327
• feat(nika-dap): admit to workspace — all 12 gates passed by @ThibautMelen in #325
• feat(hygiene): vector 9 gains counts parity — the vitrine tracks canon.yaml by @ThibautMelen in #337
• docs(privacy): the public surface drops its two studio-internal paths by @ThibautMelen in #343
• fix(changelog): the conformal cite says Theorem by @ThibautMelen in #340
• feat: nika:image_fx — deterministic artistic effects (stdlib §Media graduate #3) by @ThibautMelen in #338
• docs(privacy): the rules index drops its operator memory slugs by @ThibautMelen in #344
• feat(nika-chart): nika:chart builtin — deterministic chart artifacts (stdlib §Media #3) by @ThibautMelen in #339
• feat(provider): openrouter calls carry the app-attribution pair by @ThibautMelen in #347
• fix(cli): the run card advertises the full 64-hex chain head by @ThibautMelen in #348
• docs(readme): the model bench earns its line by @ThibautMelen in #350
• feat(nika-chart): the design pass — per-mode palettes, quantized cells, capped bars by @ThibautMelen in #349
• fix(nika-schema): the broken modeline names itself — cause, not symptom by @ThibautMelen in #351
• fix(nika-chart): the PNG heatmap speaks the SVG's quantized bins by @ThibautMelen in #355
• refactor(nika-dap): the trace forensics come home — the cli keeps the render by @ThibautMelen in #356
• feat(agents): the plugin grows its three slash commands by @ThibautMelen in #357
• fix(cli): the cost floor prices the EFFECTIVE model — and check previews it by @ThibautMelen in #352
• feat(cli): the MODELS rung — every model resolves in THIS binary by @ThibautMelen in #353
• feat(cli): wire learns opencode and hermes — the kit's two native readers by @ThibautMelen in #354
• feat(mcp): nika_check carries the MODELS rung — the third lane closes by @ThibautMelen in #358
• feat(nika-schema): egress to outputs — the workflow boundary earns its valve by @ThibautMelen in #359
• docs(changelog): the 0.99 preflight — the night's ships enter the record by @ThibautMelen in #360
• feat(run): the epilogue seam — and the run learns to say recovered and still-running by @ThibautMelen in #361
• chore(release): v0.99.0 — the audit tells the whole truth, the run narrates its own by @ThibautMelen in #362

Full Changelog: v0.98.0...v0.99.0