Changelog¶

Full release notes are available on GitHub Releases.

Note: this changelog is not exhaustive. Older releases, including the v1.x and v2.x lines, are covered by the README Version History and GitHub Releases.

v3.4 — Market + Gateway Hardening¶

v3.4.10¶

Repository alignment. The topic (fish-trail) plugins in lib/plugin/ and .opencode/plugin/ were synced to the pack copies, closing a delivery gap: since 6f7223c only the pack copies carried the cache-first context-injection rewrite and the three topic-context-filter fixes (parallel topic reads, single-topic guard correction, archive-failure abort), so installs never received them. Removed two dead artifacts aimed at the installers deleted in v3.0 (scripts/check_installer_parity.py, tests/test_migration_e2e.py); the latter is replaced by tests/test_installer_migration.py, which exercises install.py's v0.9 migration directly. Fixed verify_install.py, which expected the shelved fish-trail-compaction.ts and required 4 plugin registrations, so /petfish verify failed on every install. Corrected stale installer commands and package claims across the docs site, pack READMEs, and website; filled the v2.x and v3.x gaps in the README version history and both changelogs. pre_release_check.py now baselines pack version drift on the highest released tag instead of git describe, which resolved to a stale tag on dev. Packs: fish-trail 1.4.0, petfish-companion-skill 1.9.3, research-skill-pack 1.0.2, doc-reader-skill 1.0.1.

v3.4.9¶

fish-trail topic plugins (context injection and message filtering) are now opt-in and scoped to the context pack, registered with "enabled": false; previously they were default-on for any L1 pack and could not be disabled on OpenCode v1.18.3. The topic-aware compaction plugin was shelved after a measured null result (p=0.928). Also fixed a pre-existing syntax error in the pack copy of system-prompt-context-inject.ts and corrected public copy that still advertised the shipped compaction plugin. fish-trail 1.2.0 → 1.3.0.

v3.4.8¶

P0 fix for a companion-gateway contract violation: output.system is a shared string[] that plugins may only push to, but the gateway used +=, poisoning every later plugin's push call and silently discarding its own injected content since release. One-line fix (output.system.push()) applied across 3 copies, plus a 7th permanent release gate that statically scans all pack plugins for output field rebinding. Companion 1.9.2.

v3.4.7¶

Integrated skills.sh and ClawHub as first-class fish-market sources (13 total). skills.sh uses the official CLI search endpoint (/api/search); ClawHub is the OpenClaw registry (10.7K+ skills, documented /v1/search, 3000/min unauthenticated, includes Chinese skills). Testing hit a Chinese Gantt-chart generator via ClawHub and four dedicated skills via skills.sh. Companion 1.9.1.

v3.4.6¶

MCP Official Registry dedupe now keys on plain name, because multiple remote entries for the same server share a display name.

v3.4.5¶

MCP Official Registry results are deduplicated by name+url, since its weak q filter returned duplicate entries.

v3.4.4¶

Reworked /petfish load discovery into a marketplaces-first ladder after three root causes pushed it into minute-long GitHub mining: 4 of 8 search sources had been dead since launch (an urllib.request.quote AttributeError silently swallowed per-source), Chinese queries against English indexes returned zero results with no retry hint, and there was no enforced "marketplace before mining" step. Added ClaudSkills (69K+ SKILL.md, Chinese-capable), PulseMCP, and the official MCP Registry; Chinese zero-result queries now retry in English. Companion 1.9.0 / toolchain 0.3.1.

v3.4.3¶

Upgrade flow rewritten natural-language-first: the one-liner is the default entry and the agent runs all commands, so users stay in chat. Step 8's closing report now includes an install health matrix, current capabilities, per-platform restart guidance, and a post-restart conversational example.

v3.4.2¶

Added /petfish verify, a one-command visual verification built on verify_install.py: a 10-item PASS/FAIL checklist (plugins, registration, MCP, vault self-test, gateway markers, L1 rule repair, index quality, version comparison, commands, course capability) with per-item remediation and a gate-compatible exit code. Un-upgraded projects can run it from the raw master URL. Companion 1.8.0.

v3.4.1¶

Fixed install.py writing a legacy minimal skill-index.json, which disabled gateway domain matching in user projects after v3.4.0. The installer now emits domains (from the shipped catalog_query.py single source), per-skill pack attribution, and best-effort market available_packs. Verified live at 110 skills / 23 domains / 107 attributed / 15 market packs.

v3.4.0¶

Completed all remaining phases of dynamic skill loading (P2+P3) and the courseware upgrade (P1+P2+P3). Skill loading gained gap-aware discovery, /petfish load <name|keyword> [--install], semantic repo mining with golden-repo benchmarks, market-side trigger derivation, and vault usage tracking. Courseware gained a pedagogy reference layer plus course-assessment-design, an LLM-as-judge evals harness with CI gate, three offline teaching artifacts, and course-delivery-review. Companion 1.7.0 / toolchain 0.3.0 / course 1.5.0.

v3.3 — Dynamic Skill Loading + Courseware¶

v3.3.0¶

First milestone of dynamic skill loading (P1) and the courseware upgrade (P0). The new skill-vault MCP server delivers skill content as tool return values to bypass per-session skill discovery caching, exposing vault_index, vault_fetch, vault_stage, and vault_install; the Gateway injects a top-3 discovery block only when a capability gap has uninstalled candidates. Courseware P0 added machine-checked outline constraints, dual confirmation gates, and an authoring completion checklist. Companion 1.5.0 → 1.6.0 / course 1.3.2 → 1.4.0.

v3.2 — Skill Machinery Foundation + Installer Hardening¶

v3.2.1¶

Skill machinery foundation fixes (F1-F6): single-source triggers (skill-index.json gains a domains map, the gateway reads the index instead of a hardcoded table, and gateway_classifiers.py imports the catalog directly), --skill NAME granular install, a marketplace index packs/skills dual-key fix, project-aware suggest ranking, doc alignment, and a regression fix for calibrate eval phrases. Companion 1.5.0 / toolchain 0.2.0 / init 1.2.1. Pre-release gate 6/6 PASS.

v3.2.0¶

Fixed a systematic upgrade bug where L1_PACK_MAP mapped companion and toolchain to the same rules file, so --pack all (alphabetical) let toolchain overwrite companion's Gateway Trace rules on every upgrade; the L1 mapping was split and the stale pack copy removed. Added release gate item 6, requiring a pack-manifest version bump when pack content changes versus the latest tag, which immediately caught two existing drifts. Extended the upgrade doc with Step 3.5 delivery verification. Companion 1.3.0 → 1.4.0 / fish-trail 1.1.0 → 1.2.0.

v3.1 — Performance & Cache Architecture¶

v3.1.0¶

Multi-agent orchestration (Phases 0-5): task() spike validated (1.43x parallel speedup), skill I/O contracts (3 pilot skills), orchestration hints in companion-gateway, dispatch tracking, result aggregation with conflict detection, autonomy levels (suggest/delegate/auto). Documentation/website updated. Council-thinking references trimmed.

v3.0 — Companion Overhaul¶

v3.0.0¶

Programmatic companion-gateway.ts (6-step enforcement via TypeScript plugin). topic-context-filter fixes (placeholder bug, effective topic detection, per-topic message archiving). Legacy installers deleted (install.py sole entry). skill-index.json (100 skills). Market CLI. Web-grounding rules. 13/13 registry consolidated to monorepo. 102/102 agentskills.io compliant. 2 new packs: drawio-radar-chart, typst-pdf-builder.

v2.2 — Council Thinking + Pack Rename¶

v2.2.0¶

Pack rename anti-sycophancy-calibration-pack → judgment-calibration-pack; added council-thinking skill (5+1 multi-perspective adversarial reasoning); pack now contains 2 skills (fish-calibrate + council-thinking); alias calibrate unchanged; legacy_names preserves upgrade compatibility.

v1.9 — Testing Team Issue Resolution + Delivery Pipeline Fix¶

v0.11 — Companion Gateway Enhancement: Proactive Intelligence¶

v0.11.7¶

Documentation catchup — companion-gateway docs (EN+ZH), README, website updated to reflect 6-step Gateway flow; Token Cost Engineering blog post published.

v0.11.6¶

Companion Gateway 6-step implementation complete — all six steps (Mode Read, Topic Check, Failure Signal Detection, Skill Sense, Anti-Sycophancy Check, Proceed) integrated and operational.

v0.11.5¶

Rigor threshold refinement — only Momus plan+review for 3+ step or 3+ file tasks; simpler tasks get assumption-stating and post-verification without formal plan files.

v0.11.4¶

Anti-Sycophancy Check (Step 2.5) — rubric-first evaluation, mandatory counter-argument search before agreeing; proactivity level linked to Rigor mode (off=explicit only, on=implicit+assertions).

v0.11.3¶

Rigor Mode — rigor: true in project-mode.yaml adds plan-then-review discipline: formal plan files for complex tasks, Momus review before implementation, explicit assumption-stating. Forced on when depth: thorough.

v0.11.2¶

Project Mode (Step 0) — depth (urgent/balanced/thorough) and rigor (on/off) axes in .opencode/project-mode.yaml; session-only verbal overrides without file writes.

v0.11.1¶

Failure Signal Detection (Step 1.5) — scan previous assistant turn for known failure patterns (PDF/deploy/test/research/context), recommend matching pack if uninstalled. Implemented via catalog_query.py --check-failures.

v0.11.0¶

Gateway expansion from 3 steps to 6 steps — add Mode Read, Failure Signal Detection, and Anti-Sycophancy Check to the always-on Companion Gateway flow.

v0.10 — Research Pack Expansion: 7 Domains¶

v0.10.10¶

Auto-update capability — check_installed.py --check-updates queries GitHub latest release and compares installed pack versions; catalog_query.py --upgrade shows OS-appropriate upgrade command; Companion Gateway now checks for updates on session start; /petfish upgrade command added. Also fixes missing research alias in KNOWN_PACKS.

v0.10.9¶

Systemic trigger keyword coverage fix — align all skill descriptions with body trigger words across all 11 packs (~74 skills updated); add check_trigger_coverage() lint rule to lint_skill.py; integrate trigger-coverage into run_gate.py decision logic; add Description-Body alignment discipline to root AGENTS.md; expand research triggers in catalog_query.py. Closes #91, #89, #88.

v0.10.7–v0.10.8¶

Fix research pack integration — complete 9-touchpoint checklist for research pack (remote installer, companion catalog, README, docs, website). Crystallize "one audit, one fix" development lesson.

v0.10.6¶

Fix 4 backlog issues — replace duplicate QA script with qa_scan.py (#80), add --target to suggest for fixture isolation (#73), document JSONL/Markdown design and improve research pack UX (#79), add hybrid semantic+keyword trigger scoring with --semantic flag (#77). Closes #80, #73, #79, #77.

v0.10.5¶

Adapter skills — 4 lightweight domain adapters (travel, conference, training, content-selection) that enhance main research chains with domain-specific fields and checklists. Pack now has 50 skills.

v0.10.4¶

Risk-procurement and experience-event research domains — 11 new skills. Pack now has 46 skills.

v0.10.3¶

Learning and decision research domains — 7 new skills. Pack now has 35 skills.

v0.10.2¶

Planning research domain — 6 new skills. Pack now has 28 skills.

v0.10.1¶

SKILL_builder eradication — 10 stale refs fixed across 6 files; catalog_query.py fallback now returns actual counts. Closes #87, #86.

v0.10.0¶

Product research domain — 5 new skills (user-research, competitor-analysis, opportunity-mapper, validation-planner, decision-brief). Pack now has 22 skills.

v0.9 — Research Skill Pack¶

v0.9.6¶

Fix smoke fixture missing adr/ directory (#85); fix trigger eval runner to glob all evals/trigger/*.json (#84).

v0.9.5¶

Fix SKILL.md schema mismatches in 4 research skills (#83, #82, #78); fix repo_inventory.py node_modules inclusion (#81); fix all 4 installers writing zeroed skill/command/agent counts (#71). Closes 5 issues.

v0.9.4¶

Research pack scientific domain — 7 new skills (citation-auditor, literature-review, gap-finder, methodology-designer, experiment-planner, paper-writer, review-rebuttal). Pack now has 17 skills.

v0.9.3¶

Research pack installable — pack-manifest, installer registration, companion catalog integration, README and CHANGELOG updates.

v0.9.2¶

Research pack QA infrastructure — seeded fixtures, E2E smoke tests (15 pytest), trigger-eval harness, local smoke runner, CI gates. Closes #74, #75, #76.

v0.9.1¶

Research alias added to all 4 installers and companion catalog.

v0.9.0¶

Research skill pack MVP — 10 core skills, 7 JSON schemas, 9 Python scripts, pack infrastructure.

v0.8 — Multi-Platform & Agent Discipline¶

v0.8.1¶

Universal agent principles (cross-repo protection, network retry); complete ops AGENTS.md template; code profile experience crystallization; deployment-executor references for private repo access. Closes #66, #67, #68, #69.

v0.8.0¶

Multi-platform instruction file generation (#63) — detect_all_platforms(), content condensation for token-limited platforms, Claude Code hook scripts, uv-first Python policy enforced across project.

v0.7 — Stability & Pack Versioning¶

v0.7.2¶

Fix #57 root cause (grep -qF replacing echo | grep); fix #65 (8 missing QA bilingual terms in topic_detector.py).

v0.7.1¶

Fix #57 legacy name awareness in merge_agents_md; bump fish-trail and petfish-companion-skill to 1.0.0 (#64); fix corrupted AGENTS.md markers; update all 4 installer scripts.

v0.6 — Companion Narrative¶

v0.6.4¶

Bilingual website and docs; archive outdated v0.2 docs.

v0.6.3¶

Companion narrative rebrand; fix #57 --force upgrade bug.

v0.6.2¶

Fix companion pack skill sensing, installer dedup, catalog fallback, and universal platform detection.

v0.6.1¶

Fix topic_graph persistence, schema alignment, and intent-aware skill sensing.

v0.6.0¶

Companion Gateway with always-on topic check, 3-tier skill sense, and debug mode.

v0.5 — Fish Trail and Repo Rename¶

v0.5.4¶

Fix missing version field in topic_graph and stale detection in topic_report.

v0.5.3¶

Add the agent upgrade guide and web upgrade prompt.

v0.5.2¶

Add the v0.4.x → v0.5.x upgrade guide.

v0.5.1¶

Pre-release documentation and test suite updates.

v0.5.0¶

Rename SKILL_builder to petfish.ai; rename context-router to fish-trail; add 31 MCP tools, installer aliases, state directory migration, and topic routing scripts.

v0.4 — Context Router and Session Management¶

v0.4.10–v0.4.12¶

Add topic-aware session management with 10 new MCP tools, cross-session resume, boundary policy, activity query, agent attribution, and topic recommendations. Fix install guidance, trigger extraction scope, and deploy_dirs false positives.

v0.4.5–v0.4.9¶

Fix MCP schema issues, platform-specific restart hints, CJK detection, and trigger evaluation.

v0.4.0¶

Add the context router pack with topic detection, contamination scoring, context isolation, and 18 MCP tools.

v0.3 — Quality and Platform Hardening¶

Anti-sycophancy calibration pack.
Style v4 AI slop detection.
Release discipline with auto-resolved latest release tags.
UTF-8 fixes for PowerShell.
Comma-separated multi-pack install.

v0.2 — Skill Lifecycle Management¶

Phase 1: 8-platform adapter and companion skill with sensing, equip, and governance capabilities.
Phase 2: Marketplace search, skill authoring, and quality linting.
Phase 3: Repo mining, security audit, and publish quality gate.
Phase 4: Description optimization, trigger evaluation, and usage tracking.

v0.1 — Skill Installer¶

Multi-pack installer with remote install support.
7 skill packs: course, deploy, testdocs, petfish, ppt, init, and companion.