v1.4.1 — the mischief release

The verifier survived its own red-team.

Mischief audit — protocol-level 5/5

The verification protocol was red-teamed with five classed, seeded defects behind a sealed answer key: clean-grepping stub, wired-but-wrong, phantom commit, orphaned artifact, missing artifact. The blind verifier + runtime step caught all five (4/5 at the static ladder alone — wired-but-wrong is designed to need runtime evidence, and the protocol is the unit under test). It also flagged the fabricated "verified live" claims unprompted. Receipts: .potion/phases/09-mischief/.

Verify hardening

• verified_at is now derived mechanically (date -Iseconds), never hand-typed
• New permanent ratchet: every SKILL.md stays under the ~150-line bar (check-skill-line-bar)

Docs

• README caught up to the 1.x story: new "Knowledge that compounds" section (journal → promotion → pages → lint), brewed-itself claims moved to v1.4.0, mischief receipts cited. Repo About updated to match.

Full changelog: v1.4.0...v1.4.1

v1.4.0 — the self-audit release

Potion is a lean project framework for Claude Code — plans are prompts, disk is memory, evidence or it didn't happen.
1.4.0 completes the knowledge loop: the layer now audits itself — four hunts, automatic cache repair, and a hard human gate on ground truth.
Proven live with a seeded-defect kill and hash-verified untouched journals — receipts in .potion/phases/08-lint/.

1.4.0 — 2026-07-05

The self-audit release: the knowledge layer now lints itself — cache fixes are
automatic, truth changes are human-gated. This completes the Karpathy-pattern
set: journal (append-only truth) → pages (distilled cache) → lint (the
auditor that keeps them honest).

Lint — four hunts over the knowledge tree

• New "Lint" section in the learn skill, run over project scope then user
  scope (commit fe6c069): (1) contradictions — semantic judgment of entry
  insights and page syntheses, Claude is the linter; (2) stale claims —
  files backpointers gone from disk, page claims superseded by newer
  same-key entries; (3) orphans — page ### {key} blocks with no live
  journal line; (4) coverage gaps — confidence >= 8 live entries no page
  distills
• Trigger: runs automatically after every Distill, or on demand via
  /potion:learn lint
• The safety line — cache-auto, truth-human: cache-side findings (orphans,
  stale page claims, coverage gaps) are fixed by rebuilding the affected
  page whole from the journal; truth-side findings (contradictions, any
  urge to tombstone or reword a journal entry) are NEVER self-resolved
• Parked routing: truth-side findings go to STATE.md ## Parked, one line,
  both sources cited — the next /potion:discuss tables them
• CORE contract: lint safety line under Learnings, core/CORE.md
  (commit a8f7973)

Evidence

• Live proof on the real knowledge tree, both scopes: clean baseline lint;
  a transparently seeded cache-side orphan detected by hunt 3 and
  auto-rebuilt away (rebuilt page byte-identical to HEAD); truth journals
  sha1-identical across the entire run — the linter has teeth and a muzzle:
  .potion/phases/08-lint/evidence/02-lint-transcript.txt (commit d7b24cb)

v1.3.0 — the distillation release

Potion is a lean project framework for Claude Code — plans are prompts, disk is memory, evidence or it didn't happen.
1.3.0 adds the distillation layer: journals synthesize into topical, source-backed knowledge pages — fragments become compounding understanding.
The repo's own .potion/knowledge/ is the live showcase; receipts in .potion/phases/07-distill/.

1.3.0 — 2026-07-05

The distillation release: journals stay ground truth, pages make them readable.

Distillation pages

• New pages layer in two scopes: project journal distills to
  .potion/knowledge/ (committed with the repo), user journal to
  ~/.claude/potion/pages/ (outside git) — each with an index.md listing
  every page with its one-line scope and entry count
• Pages are a regenerable cache, never ground truth: rebuilt whole from the
  deduped journal, never hand-edited, safe to delete; when a page and the
  journal disagree, the journal wins and the page is due a rebuild
• Source-backed claims: every ### {key} block cites the journal entry key
  it distills (user scope also cites the source repo) — an uncited claim is
  a defect
• CORE contract: "Distillation pages" subsection under Learnings,
  core/CORE.md (commit 55b4fa7)
• Procedure: new "Distill" section in the learn skill — dedup newest-wins,
  drop tombstones, cluster 2-6 entries per page, regenerate pages whole,
  rewrite index (commit a641ecb)
• Planners prefer synthesis: planning reads pages before raw journal
  fragments when both exist (commit f509289)
• Trigger honesty: automatic distillation fires at compaction (~500 lines),
  which no journal has reached — the trigger is dormant until then. On-demand
  runs via /potion:learn distill work today and produced the evidence below.

Evidence

• First real distillation, both scopes: 9 project entries → 3 pages + index
  in .potion/knowledge/; 3 user entries → 1 page + index in
  ~/.claude/potion/pages/; every citation spot-checked to exactly one
  journal line:
  .potion/phases/07-distill/evidence/02-distill-transcript.txt
  (commit 371fc19)

v1.2.0 — the harvest release

Potion is a lean project framework for Claude Code — plans are prompts, disk is memory, evidence or it didn't happen.
1.2.0 closes the loop on knowledge itself: every passing phase is harvested before its verdict is sealed, and the best of it flows to the cross-repo journal from 1.1.0.
Proven retroactively on potion's own 119-agent audit phase — receipts in .potion/phases/06-ingest/.

1.2.0 — 2026-07-05

The harvest release: no closed phase goes cold with its insights unharvested.

Phase-close harvest

• /potion:verify step 4 gains "Harvest before verdict" (pass verdicts only):
  sweep the phase's SUMMARY Deviations/Concerns, the verifier's
  ladder/panel reasoning in VERIFICATION.md, and any .potion/debug/ files
  closed during the phase; distill what clears the 5-minute bar into
  .potion/learnings.jsonl — max 5 entries, files backpointers to the
  source artifact (commit f5e260b)
• Duplicate guard: grep existing keys first, harvest only what fell through
  during the phase, never re-log; zero new entries is a valid outcome, stated
  explicitly
• CORE principle: "Phase-close ingest: a pass verdict triggers a harvest of
  the phase's SUMMARY/VERIFICATION/debug reasoning into learnings" —
  Learnings section, core/CORE.md (commit e3b1438)
• Composes with 1.1.0's promote-up bar: harvested entries ride the normal
  CORE Cross-repo knowledge bar (confidence >= 8, generalizable, loop-born)
  into ~/.claude/potion/knowledge.jsonl

Evidence

• The procedure run for real against the cold storage that motivated it —
  01-harden's 42-finding audit reasoning: 9 artifacts swept, duplicates
  skipped, 3 distilled entries appended, 2 promoted up:
  .potion/phases/06-ingest/evidence/02-retro-ingest-transcript.txt
  (commit 713dc61)

v1.1.0 — the compounding release

Potion is a lean project framework for Claude Code — plans are prompts, disk is memory, evidence or it didn't happen.
1.1.0 adds the cross-repo knowledge layer: what one repo learns, every repo knows.
Built and verified by potion's own loop — receipts in .potion/phases/05-compound/.

1.1.0 — 2026-07-05

The compounding release: what one repo learns, every repo knows.

Cross-repo knowledge journal

• New user-scope journal ~/.claude/potion/knowledge.jsonl — same grammar as
  .potion/learnings.jsonl plus a mandatory provenance field:
  {"source":{"repo":"<repo-name>","key":"<original-key>","ts":"<date>"}}
• Promotion bar (CORE): confidence >= 8, generalizable beyond the source repo
  (tools, frameworks, protocols — never business logic, secrets, or private
  URLs), and born inside potion's verified loop
• Memory-poisoning guard: external text is never promoted — only loop-born
  entries cross the repo boundary
• Append-only, newest-wins by key; interleaved appends from parallel sessions
  absorbed by that rule; same ~500-line compaction as the repo journal
• /potion:learn promotes up when the bar clears ("Promoted: {key}"), creates
  the directory on first use; when in doubt, don't

Delivery

• Planners grep ~/.claude/potion/knowledge.jsonl in plan step 2 and cite
  matches as Prior cross-repo learning: {key} (from {repo})
• session-start hook surfaces the newest journal entries under a
  "Cross-repo knowledge" header with (from {repo}) provenance

Evidence

• Real end-to-end run — genuine promotion, live hook delivery, plan-side
  grep + attribution: .potion/phases/05-compound/evidence/03-e2e-transcript.txt

v1.0.0 — the campaign release

Potion is a lean project framework for Claude Code: eleven skills, two agents, one hook.
The filesystem is the memory, plans are prompts, and nothing is "done" without evidence.
This release was built by potion running on itself — the full campaign record is committed at .potion/.

1.0.0 — 2026-07-05

The campaign release: potion brewed itself. .potion/ in this repo is the
live record — design, three phases, two first-cycle pass verdicts so far
(01-harden and 02-empower, both cycle 1, zero gaps). The audits and fixes
below were executed by potion's own loop: workers, blind verifiers, refuter
panels.

01-harden — six-lens audit

• 119 agents over 27 files (10 finders per lens, 3-refuter adversarial
  panels, refute-by-default), full coverage, no finders lost
• 60 raw findings → 42 unique; 25 confirmed by the panels; 11 refuted
• 20 behavioral defects fixed across 6 plans / 14 commits; 5 style findings
  accepted with reasons on record. Highlights:
  • Verifier never emits VERIFIED — the orchestrator is the single promotion
    point
  • Ratchet checks take their own check-{source} key, so newest-wins dedup
    can no longer shadow a lock
  • init gained a real new-cycle procedure (and PROJECT.md its ## Phases
    section)
  • Lossless pause↔resume: closed-vocab status captured and restored
  • Gap acceptance is user-only
  • One law canon
  • Install modes that work (skills-dir mode removed)

02-empower — six mechanisms, every one citing dogfood evidence (Law 10)

• evidence/ convention — canonical home for runtime proof artifacts,
  referenced by path from VERIFICATION.md
• Check-runner contract — POSIX sh from repo root, trimmed-stdout vs exit N
  semantics; kills the || true hack
• Build-freshness precondition — verify confirms the running app corresponds
  to HEAD before judging behavior
• Learnings delivery — planners grep learnings.jsonl into TECHNICAL NOTES;
  worker digests carry matching entry keys
• Concern→Parked routing — non-gap findings land in STATE.md's Parked list;
  nothing evaporates
• Continuity freshness — any skill that moves position refreshes the
  Session-continuity block

03-polish

• README truth-checked against HEAD and given a stranger-ready
  "Quickstart: your first brew"
• This release