Quartz 5

changelog

11 min read

Algorithm Changelog

Each entry: version, date, what changed, provenance, migration/rollback notes.

v5.7.11 — 2026-06-18

ISA Troubleshooting Log and cross-engine troubleshooting protocol. Additive — all gates, rules, and phases preserved verbatim from v5.7.10; only the ISA body section listing, a new Troubleshooting Method section, and the preamble changed.

  • ISA ## Troubleshooting Log body section added. The ISA body now recognizes ## Troubleshooting Log as a fifth section (alongside Context, Criteria, Decisions, Verification). Required at E2+ when Gate A fires (debugging). Captures the iteration chain: probe → hypothesis → fix → re-probe for each attempt. Format per PAI/DOCUMENTATION/TROUBLESHOOTING_METHOD.md. E1/additive work skips it.
  • Cross-engine troubleshooting protocol referenced. Algorithm now names PAI/DOCUMENTATION/TROUBLESHOOTING_METHOD.md as the canonical troubleshooting protocol. The 5-step loop (show failing state → state hypothesis → try one fix → re-probe → iterate) is defined there; the Algorithm defines when the log is required.
  • Engine configs updated in same change: PNC settings.json verbose: false → true; PNX .codex/AGENTS.md gains troubleshooting section + --verbose default; PNK .opencode/AGENTS.md gains protocol reference; PNC CLAUDE.md gains cross-engine references section.

Provenance: Duane asked “how can we enable the other PAI engines to follow the method of thought and troubleshooting that opencode displays” (2026-06-18). Cross-engine gap analysis produced the three-layer solution (protocol doc + engine configs + Algorithm ISA log). Duane: “Proceed” on adding the Troubleshooting Log to the Algorithm. Migration: none (additive doctrine + additive engine config edits). Rollback: set LATEST → v5.7.10; revert engine config edits.

Mined candidates #1 + #3 from the 2026-06-16 reflection-mining run. Additive — all gates/rules/phases preserved verbatim from v5.7.9; only the Gate A row and a new BUILD checkpoint changed.

  • OBSERVE Gate A extended to infra-attribution (#1, dominant 20+-hit theme). For any root-cause/attribution claim about a host/ACL/policy/mount/timing, probe the authoritative dynamic source (read the ACL policy, measure host load, pai-infra.ts resolve) before forming the hypothesis. Implemented as a Gate A extension, not a 7th gate (v5.7.9 advisor anti-proliferation guidance); cites-and-extends CLAUDE.md “resolve, don’t recall” + feedback_verify-gated-conditions-before-attribution rather than restating. Distinct from Gate F (precondition existence). Provenance: guessed-ACL-from-ping-timeout; model-speed conclusion before measuring the load-34 host storm.
  • BUILD Backend-Falsification Checkpoint (#3, the s5 over-budget run). Before building a pipeline that writes to a stateful/external backend, throwaway write+read-back on a disposable key first — confirm the write method persists. Provenance: a full auto-on-close profile-writer built before discovering state.vscdb is unreachable by file edits.

Companion same-selection changes outside this file: #2 — deadman-registry.json gains an omnipulse-digest-distribution total-silence probe + an _outcome_doc (liveness-of-trigger ≠ liveness-of-outcome; per-channel outcome-liveness deferred as named follow-up — no failure-only artifact exists yet). #4 — PAI_SYSTEM_PROMPT.md Operational Rule: grep all callers before claiming a shared-contract fix complete.

Provenance: Duane selected all four mined candidates (review ISA 20260616-085146; mining via PAIUpgrade; build ISA 20260616-092545_apply-4-mined-candidates). Migration: none (additive doctrine + additive registry/constitution edits). Rollback: set LATEST → v5.7.9; revert the deadman-registry.json and PAI_SYSTEM_PROMPT.md edits via their checkpoint SHAs.

v5.7.9 — 2026-06-16

Design-time prior-art consult promoted into OBSERVE. Additive — all gates/rules/phases preserved verbatim from v5.7.8; only OBSERVE Step 0 (new item 3) and a one-line PLAN cross-ref changed.

  • OBSERVE Step 0 item (3): consult prior art + probe schema in OBSERVE, not later. Mined from the 2026-06-16 reflection review — 12/55 unmined reflections (≈22%, the single dominant theme) cited a prior-art / feedback-memory / schema read that landed at PLAN/THINK/BUILD when it belonged in OBSERVE. Unconditional prior-art/root-cause consult (MEMORY/KNOWLEDGE/ + feedback_*.md for design-bearing hits); conditional schema/tool-grammar probe (only when an ISC depends on one); relevance framed as search-then-consult. Differentiated from PLAN’s FEEDBACK MEMORY AUTO-CONSULT by purpose (OBSERVE = design-time prior-art; PLAN = pre-action gotcha re-check on the chosen approach), with reciprocal cross-references so the boundary survives a real session. Harm anchors: a wrong NetworkManager DNS root cause the advisor had to overturn (feedback memory read at PLAN not OBSERVE); an unmeasurable FP-gate claim made before probing the stripped-path telemetry schema. Falsification: the next /pai-upgrade mine-reflections run measures recurrence/decay of this theme. Advisor verdict on the change: concerns (boundary under-specification + missing falsification loop) — both addressed before write. Provenance: Duane — “address findings” on the reflection review (sibling ISA 20260616-085146_review-reflections; build ISA 20260616-085748_observe-prior-art-consult-v579). Migration: none (additive doctrine). Rollback: set LATEST → v5.7.8.

v5.7.8 — 2026-06-14

Self-identifying voice announcements. Additive — only voice-announcement wording changed; all gates/rules/phases preserved verbatim from v5.7.7.

  • Every Algorithm voice announcement states the task. Entry is "Entering the Algorithm for <the 8-word TASK>" and each phase transition is "Entering the <Phase> phase for <the TASK>" (from the 🗒️ TASK banner) instead of context-free phrasing; the bare phrasing is forbidden across entry and all seven phases (OBSERVE→LEARN). Provenance: Duane — “announcement for entering the algorithm should state what for” + “we need the same for all of the algorithm steps” (2026-06-14). Migration: none (wording only). Rollback: set LATEST → v5.7.7.
  • Rapid-phase coalescing (runtime, VoiceServer/server.ts). Trailing-debounce: when phase transitions fire within ~5s of each other, only the last keeps the spoken task context (earlier ones bare). Agent always emits the full form; coalescing is VoiceServer-side. Provenance: Duane — “if the phases fall within 5 seconds of each other, only add the context to the last one.”

v5.7.7 — 2026-06-13

One additive block from the same 2026-06-10 mining run that produced Gate F + v5.7.6 (candidate #5). Byte-copy of v5.7.6 + title version + the Quota-Check Gate; every prior rule preserved.

  • PLAN-phase Quota-Check Gate (mined #5, 7/272 ≈ 3%): before spawning ≥2 parallel agents OR any Agent call with model:"opus", probe subscription headroom via bun PAI/Tools/UsageProbe.ts. Three-tier action (proceed / downshift / defer); inserted in PLAN after the COORDINATION GATE. Pairs structurally with the Local-First Routing Policy in CLAUDE.md — that policy says prefer local; this gate makes “before heavy spawn, check” the procedural step.

Provenance: cited 2026-03-06 reflection where the executor spawned expensive parallel agents and the user manually interrupted to redirect to Ollama. Migration: none. Rollback: set LATEST back to v5.7.6.

v5.7.6 — 2026-06-12

Two additive blocks from the same 2026-06-10 mining run that produced Gate F (candidates #2 and #3). Byte-copy of v5.7.5 + title version + the two blocks; every prior rule (incl. Gate F) preserved.

  • OBSERVE Step 0 — Batch & Parallelize Inputs (mined #2, 24/272 ≈ 9%): read all known target files together and launch independent background jobs in one parallel round before deep reads. Inserted in OBSERVE after the preflight output block.
  • Persistent-service completion gate (mined #3): VERIFY Rule 1 artifact table gains a row — network-listener / long-running-service ISCs require a running AND enabled unit (systemctl is-active + is-enabled), never a foreground/test process.

Provenance: /pai-upgrade mine-reflections over 272 reflections (2026-06-10). Migration: none (additive). Rollback: set PAI/Algorithm/LATEST back to v5.7.5.

v5.7.5 — 2026-06-10

OBSERVE Preflight Gate F (Prerequisite Probe) added — additive; byte-copied from v5.7.4 with only the title version line + the Gate F table row changed. For build/edit/deploy ISCs that write a path, start a service, or call an external tool/host, probe the precondition empirically (writability+ownership, tool-on-PATH, host/mount reachability, port/process conflict) before committing the ISC.

Provenance: /pai-upgrade mine-reflections over 272 algorithm reflections (2026-06-10) — Gate-F’s theme was the single dominant pattern at 96/272 (35%, still 39% in recent runs). Canonical live case: the 2026-06-08 backup-infra session (NFS permission wall, 13-day-silent qdrant backup, three path-existence bugs — all Gate-F catches). The mining itself was only possible after fixing the reflection-pipeline PAI_DIR split (commit fcd2e66) that had left the store unreadable.

Migration: none (additive). Rollback: set PAI/Algorithm/LATEST back to v5.7.4.

Addition — Gate F (Prerequisite Probe)

Inserted after Gate E in the OBSERVE preflight table. Complements Gate E (prior-art existence) and Gate B (credentials) without overlap — F is runtime state; absorbs the recurring permissions-gate pattern as sub-clause (a). Fires at all tiers including E1; a blank 🚦 F: block when a write/start/call ISC exists is a doctrine violation.

v5.7.4 — 2026-05-30

One additive security-escalation clause in the UNDER-CLAIM CHECK; byte-copied from v5.7.3 (only the title version line changed). Provenance: ISA 20260530-122027_source-citation-doctrine-v574 (11-model UnifiedCouncil honesty red-team council-20260530120336-f6234f).

Addition — Security escalation (UNDER-CLAIM CHECK)

What changed: For safety-asserting claims (asserting the absence of a problem — “no vulns”, “patched”, “redacted”, “enforced”, “safe to ship”), the Honesty floor’s silent-drop exit is REMOVED: an uncited safety claim is UNVERIFIED + an open finding; the absence of a finding in a tasked domain is itself UNVERIFIED unless the all-clear check is cited (anti-omission); resolution requires a citation, not a narrative. Provenance: the council (GLM/MiniMax) found v5.7.3’s “just under-claim” wording prompt-injection-weaponizable (steer the model to silently suppress a true finding); DeepSeek/Kimi proposed the source-citation fix; advisor-hardened (anti-omission limb, directional scope, cited-resolution, softened absolute). Override: overrides “Confidence requires source” for the safety-asserting subclass only (removes the silent-drop option there).

Migration/rollback: Additive. LATEST flips to v5.7.4 after the run; rollback = restore LATEST to v5.7.3 (v5.7.4.md stays inert unless LATEST points to it).

v5.7.3 — 2026-05-30

Three additive rules promoted from battle-tested Spine patterns; byte-copied from v5.7.2 (zero rule deletions). Provenance: ISA 20260530-034955_spine-council-recommendations (Spine council post-session review).

Addition 1 — Independent Cross-Check (VERIFY, Rule 1b)

What changed: New VERIFY sub-rule. When a step extracts/derives a set from sources, run two genuinely-independent implementations and assert set identity (not just count). Provenance: Spine hooks-extract.ts A-vs-B cross-check (PATH A structured parse vs PATH B raw regex) — caught a real UserPromptSubmit under-model. Trigger: any VERIFY step where a wrong set would pass an ordinary live-probe.

Addition 2 — Under-claim deliberately (THINK design-review question)

What changed: New THINK design-review question. For every claim a component makes about its own behavior, name the mechanical proof; if you can’t, under-claim. Provenance: Spine probe.ts (“never prints enforced/verified”). Trigger: building validators/probes/guards/registries/health surfaces.

Addition 3 — E5 Session-Scope Checkpoint (conditional, event-triggered)

What changed: New conditional gate, E5 only. At a phase boundary, IF context budget high OR phase crossed a large tool-call count, ask once whether the next phase should wait for a fresh session; write a one-line Decisions note ONLY when the answer is “wait” (silent otherwise). Provenance: Spine E5 run 20260529-182417 (P3-inc2 deferral). Advisor-shaped to be event-triggered + silent-on-no, not a fixed cadence — avoids ceremony tax.

Migration/rollback: Additive only. Algorithm/LATEST flips to v5.7.3 after the run’s Cato gate passes (cutover-last). Rollback = restore LATEST to v5.7.2; v5.7.3.md can remain (inert unless LATEST points to it).

v5.7.1 — 2026-05-09

Patch 1 — Preflight Gate E (Existence Probe)

What changed: Added a fifth row to the OBSERVE preflight gates table.

Trigger: Build / add / create / implement verbs.

Goal: Run a single Glob or Grep against the natural target directory before THINK commits. Either find prior art (incorporate into ISCs) or confirm greenfield (document confirmed-greenfield: <reason> in ## Decisions).

Tier gate: ALL tiers including E1.

Provenance: 3× recurring reflection pattern across hook-observability-schema, memory-acl, audit-gaps (2026-05-03). Existing four gates (A/B/C/D) all reported N/A on greenfield-on-existing-scaffold and hardening work, leaving the most-needed coverage silent.

Verification target: Reflection mining 30 days post-release should show ≥50% drop in “should have probed earlier in OBSERVE” Q1 entries.

Rollback: Delete the new table row in v5.7.1.md preflight gates section. Doctrine reverts to v5.7.0 four-gate behavior.

Patch 2 — Rule 2 PLAN→BUILD Hard-Binding

What changed: Rule 2 commitment-boundary advisor call upgraded from soft guidance to a phase-transition gate. At E2+ effort, the voice announcement "Entering the Build phase" MUST NOT fire until advisor() returns. Added a 🛡️ PLAN→BUILD ADVISOR GATE block to PLAN phase as the operational surface.

Skip-clause hardening: Skip is invalid if PLAN produced a Deliverable Manifest with 2+ entries OR ISC count crossed the E2 floor (16). Wall-clock + file-count alone (v5.7.0) missed cases where PLAN named 3 deliverables but work happened to fit in 4 minutes. Manifest size and ISC count are PLAN-phase signals — they predict commitment scope before BUILD begins, which is the right input for a PLAN→BUILD gate.

Tier gate: E2+ (E1 fast-path skips Rule 2 entirely).

Provenance: 2× direct reflection pattern across cross-engine-memory-v0 (2026-05-01) and memory-acl (2026-05-03). Prior soft “before BUILD begins” wording was empirically being deferred until VERIFY, by which point rework cost was high.

Verification target: Reflection mining 30 days post-release should show ≥50% drop in “should have invoked advisor earlier” Q1/Q2 entries.

Rollback: Restore the v5.7.0 Rule 2 wording in ## Verification Doctrine and remove the 🛡️ PLAN→BUILD ADVISOR GATE block from PLAN phase.

Migration Notes

  • No existing ISA needs migration. The new gate fires on new OBSERVE runs only.
  • Existing in-flight Extended+ ISAs that have already passed PLAN may proceed without the new gate — applies to v5.7.1-onward sessions.
  • Advisor call cost: ~30s wall-clock per PLAN→BUILD transition at E2+. Budget impact: negligible at E3+, noticeable at E2 (~3min budget).

Files Changed

  • PAI/Algorithm/v5.7.1.md (new — copy of v5.7.0 with two patches applied)
  • PAI/Algorithm/LATEST (flipped from v5.7.0 to v5.7.1)
  • PAI/Algorithm/changelog.md (this file — created fresh; v5.7.0.md mentions a changelog but none existed on disk before now)

Approval Trail

  • Surfaced: 2026-05-08 via triple PAIUpgrade run (mine reflections + full upgrade)
  • Proposed: 2026-05-08 in MEMORY/WORK/20260508-232247_address-upgrade-items/ALGORITHM_V5_7_1_PROPOSAL.md
  • Approved by Duane: 2026-05-09 (“apply now”)
  • Applied: 2026-05-09

Graph View