Skip to content

Latest commit

 

History

History
148 lines (111 loc) · 34.4 KB

EXTENSIONS_CATALOG.md

File metadata and controls

148 lines (111 loc) · 34.4 KB

HelixCode Extensions Catalog

Revision 1
Created 2026-09-23
Last modified 2026-09-23T16:57:04Z
Status active
Status summary —
Continuation docs/CONTINUATION.md

Per-§11.4.228 (Cross-agent extension lifecycle mandate) catalog of EVERY skill registered under scripts/skills_manifest.tsv — the single T-P1.04.2 manifest scripts/register_skills.sh provisions from — with a per-platform compatibility matrix across the five governed CLI-agent platforms this constitution names as its minimum: Claude Code, OpenCode, HelixCode, Gemini CLI, Qwen Code. Generated FROM the manifest by helix_code/scripts/extensions_catalog/generate_catalog.py — never hand-transcribed — with a real, live filesystem probe per skill × platform cell, never a blank cell (§11.4.201(6)).

Closed status vocabulary (§11.4.228(A)(1)): SUPPORTED (verified working, real per-platform liveness probe this run) / CONFIG-ONLY (declared — this platform HAS a registrar targeting it — but not currently wired) / UNTESTED (this platform has NO registrar/wiring mechanism reaching this skill at all — an honest §11.4.197-tracked gap) / UNSUPPORTED (proven structurally impossible per §11.4.112 — unused by this generator; no platform here has been proven incapable of loading a Markdown-frontmatter skill).

HelixCode column — three distinct facts, not one (HXC-159 T-P9.01 round-4 F-22(b) correction). The HelixCode cell's SUPPORTED / CONFIG-ONLY status still reports the SAME thing it always has — whether the skill is reachable via .claude/skills/ by some tier (a directory-presence probe). That status alone does NOT establish either of two further, independently live-probed facts, each also-true and each distinct: (1) whether the skill is specifically ADMITTED by the external tier's own similarity-conflict policy gate (live-probed via submodules/helix_agent/cmd/similarity-probe, the SAME code path Load() uses); and (2) — for a skill the gate DOES admit — whether the FINAL registry a user actually gets skills from serves it FROM the external tier or from a same-named, independently-loaded builtin copy that overrides it (live-probed via a real helixcode skills list, HXC-159 F-22(a)'s root cause: loadBuiltinSkills runs on every load, last). Every Per-skill-detail note below states ALL applicable facts explicitly and honestly — never hand-transcribed, §11.4.6 — including when either probe itself could not run this run, in which case the note says so rather than silently reusing a stale or assumed verdict.

Table of contents

Compatibility matrix

Skill Claude Code OpenCode HelixCode Gemini CLI Qwen Code
action-prefix-system SUPPORTED (.claude/skills/action-prefix-system) UNTESTED (N/A) SUPPORTED (.claude/skills/action-prefix-system (external tier)) UNTESTED (N/A) UNTESTED (N/A)
media-validator SUPPORTED (.claude/skills/media-validator) UNTESTED (N/A) SUPPORTED (.claude/skills/media-validator (external tier)) UNTESTED (N/A) UNTESTED (N/A)
multitrack SUPPORTED (.claude/skills/multitrack) UNTESTED (N/A) SUPPORTED (.claude/skills/multitrack (external tier)) UNTESTED (N/A) UNTESTED (N/A)
reporting-workable-items SUPPORTED (.claude/skills/reporting-workable-items) UNTESTED (N/A) SUPPORTED (.claude/skills/reporting-workable-items (external tier)) UNTESTED (N/A) UNTESTED (N/A)
scheduled-work-queue SUPPORTED (.claude/skills/scheduled-work-queue) UNTESTED (N/A) SUPPORTED (.claude/skills/scheduled-work-queue (external tier)) UNTESTED (N/A) UNTESTED (N/A)
session-sync SUPPORTED (.claude/skills/session-sync) UNTESTED (N/A) SUPPORTED (.claude/skills/session-sync (external tier)) UNTESTED (N/A) UNTESTED (N/A)
skill-catalog SUPPORTED (.claude/skills/skill-catalog) UNTESTED (N/A) SUPPORTED (.claude/skills/skill-catalog (external tier)) UNTESTED (N/A) UNTESTED (N/A)
workable-item-lifecycle SUPPORTED (.claude/skills/workable-item-lifecycle) UNTESTED (N/A) SUPPORTED (.claude/skills/workable-item-lifecycle (external tier)) UNTESTED (N/A) UNTESTED (N/A)

Per-skill detail

action-prefix-system

  • Owner (registrar): constitution
  • Manifest source: constitution/skills/action-prefix-system
Platform Status Loading path Evidence / Note
Claude Code SUPPORTED .claude/skills/action-prefix-system symlink, SKILL.md read 7332 bytes (resolves to the manifest-declared source). Live-probed 2026-09-23T16:57:04Z.
OpenCode UNTESTED N/A No .opencode/skills/ directory exists in this checkout and no registrar (scripts/register_skills.sh only ever writes .claude/skills/) targets one — zero recorded loading path this cycle, tracked per §11.4.197. .opencode/ here carries only SpecKit commands (.opencode/commands/), not skills.
HelixCode SUPPORTED .claude/skills/action-prefix-system (external tier) ADMITTED by the external-tier similarity-conflict policy gate (live-probed this run via submodules/helix_agent/cmd/similarity-probe against the real .claude/skills/, qualified identity 'Action Prefix System' -- not excluded on a similarity conflict). HOWEVER the FINAL registry SOURCE (live-probed this run via a real helixcode skills list) is 'builtin', NOT the external tier -- a same-named builtin copy loaded LAST overrides the admitted external-tier entry (F-22(a)); content is byte-identical post-F-13 so this is not user-visible-wrong, but claiming this skill 'resolves via the external tier' would be FALSE. helix_code/internal/agent/skill_dirs.go ExternalSkillsDir() defaults to '.claude/skills' (config-injectable via HELIXCODE_EXTERNAL_SKILLS_DIR); the external tier's LOADING MECHANISM (a single fixture skill, no similarity conflict) is REAL-BINARY verified at qa-results/hxc159/helixcode_wire/20260922T190852Z/ (green_cli_e2e.txt). HXC-159 F-17 (2026-09-23): the shared SimilarityThreshold was recalibrated 0.0256->0.0709 against the real, frontmatter-complete corpus-1 (see docs/qa/hxc159_f17_remediation_20260923T0921Z/threshold.json); the REAL 8-9-skill corpus-1 tier is nonetheless STILL policy-refused at BOTH consumers on a genuine confusable pair (action-prefix-system || reporting-workable-items, score 0.1045 -- real content overlap, not a miscalibration), captured at docs/qa/hxc159_f17_remediation_20260923T0921Z/helixcode_recalibrated_run.txt and the mirrored helix_agent_native_recalibrated_run.txt. This SUPPORTED cell's status token reports per-skill FILE PRESENCE at the platform directory (§11.4.228(A)(1)'s liveness probe), not whole-tier ACTIVATION -- as of T-P9.01 round-4 F-22(b) (2026-09-23) this note is PREFIXED, above, by a SEPARATE, live-probed verdict on the distinct fact of external-tier-SPECIFIC policy-gate admission (see external_tier_admission in catalog.json), closing the residual gap this paragraph used to describe as unclassified. Honest boundary (§11.4.6, T-P5.05.4 in specs/001-helixskills-incorporation/tasks.md): the per-skill activation-allowlist/ceiling mechanism was NOT verifiable from within helix_code/'s own file scope this cycle — grep-confirmed zero Go implementation of a per-skill activation ceiling in this module.
Gemini CLI UNTESTED N/A No .gemini/skills/ directory exists in this checkout (only .gemini/commands/, unrelated to the Agent-Skill format) and no registrar targets one — zero recorded loading path this cycle, tracked per §11.4.197 (the exact G1 gap the constitution's own §11.4.228 forensic anchor names for this platform).
Qwen Code UNTESTED N/A No .qwen/skills/ directory exists in this checkout (only .qwen/commands/ + .qwen/settings.json, unrelated to the Agent-Skill format) and no registrar targets one — zero recorded loading path this cycle, tracked per §11.4.197 (the exact G1 gap the constitution's own §11.4.228 forensic anchor names for this platform).

media-validator

  • Owner (registrar): constitution
  • Manifest source: constitution/skills/media-validator
Platform Status Loading path Evidence / Note
Claude Code SUPPORTED .claude/skills/media-validator symlink, SKILL.md read 2225 bytes (resolves to the manifest-declared source). Live-probed 2026-09-23T16:57:04Z.
OpenCode UNTESTED N/A No .opencode/skills/ directory exists in this checkout and no registrar (scripts/register_skills.sh only ever writes .claude/skills/) targets one — zero recorded loading path this cycle, tracked per §11.4.197. .opencode/ here carries only SpecKit commands (.opencode/commands/), not skills.
HelixCode SUPPORTED .claude/skills/media-validator (external tier) ADMITTED by the external-tier similarity-conflict policy gate (live-probed this run via submodules/helix_agent/cmd/similarity-probe against the real .claude/skills/, qualified identity 'media-validator' -- not excluded on a similarity conflict). HOWEVER the FINAL registry SOURCE (live-probed this run via a real helixcode skills list) is 'builtin', NOT the external tier -- a same-named builtin copy loaded LAST overrides the admitted external-tier entry (F-22(a)); content is byte-identical post-F-13 so this is not user-visible-wrong, but claiming this skill 'resolves via the external tier' would be FALSE. helix_code/internal/agent/skill_dirs.go ExternalSkillsDir() defaults to '.claude/skills' (config-injectable via HELIXCODE_EXTERNAL_SKILLS_DIR); the external tier's LOADING MECHANISM (a single fixture skill, no similarity conflict) is REAL-BINARY verified at qa-results/hxc159/helixcode_wire/20260922T190852Z/ (green_cli_e2e.txt). HXC-159 F-17 (2026-09-23): the shared SimilarityThreshold was recalibrated 0.0256->0.0709 against the real, frontmatter-complete corpus-1 (see docs/qa/hxc159_f17_remediation_20260923T0921Z/threshold.json); the REAL 8-9-skill corpus-1 tier is nonetheless STILL policy-refused at BOTH consumers on a genuine confusable pair (action-prefix-system || reporting-workable-items, score 0.1045 -- real content overlap, not a miscalibration), captured at docs/qa/hxc159_f17_remediation_20260923T0921Z/helixcode_recalibrated_run.txt and the mirrored helix_agent_native_recalibrated_run.txt. This SUPPORTED cell's status token reports per-skill FILE PRESENCE at the platform directory (§11.4.228(A)(1)'s liveness probe), not whole-tier ACTIVATION -- as of T-P9.01 round-4 F-22(b) (2026-09-23) this note is PREFIXED, above, by a SEPARATE, live-probed verdict on the distinct fact of external-tier-SPECIFIC policy-gate admission (see external_tier_admission in catalog.json), closing the residual gap this paragraph used to describe as unclassified. Honest boundary (§11.4.6, T-P5.05.4 in specs/001-helixskills-incorporation/tasks.md): the per-skill activation-allowlist/ceiling mechanism was NOT verifiable from within helix_code/'s own file scope this cycle — grep-confirmed zero Go implementation of a per-skill activation ceiling in this module.
Gemini CLI UNTESTED N/A No .gemini/skills/ directory exists in this checkout (only .gemini/commands/, unrelated to the Agent-Skill format) and no registrar targets one — zero recorded loading path this cycle, tracked per §11.4.197 (the exact G1 gap the constitution's own §11.4.228 forensic anchor names for this platform).
Qwen Code UNTESTED N/A No .qwen/skills/ directory exists in this checkout (only .qwen/commands/ + .qwen/settings.json, unrelated to the Agent-Skill format) and no registrar targets one — zero recorded loading path this cycle, tracked per §11.4.197 (the exact G1 gap the constitution's own §11.4.228 forensic anchor names for this platform).

multitrack

  • Owner (registrar): constitution
  • Manifest source: constitution/skills/multitrack
Platform Status Loading path Evidence / Note
Claude Code SUPPORTED .claude/skills/multitrack symlink, SKILL.md read 2543 bytes (resolves to the manifest-declared source). Live-probed 2026-09-23T16:57:04Z.
OpenCode UNTESTED N/A No .opencode/skills/ directory exists in this checkout and no registrar (scripts/register_skills.sh only ever writes .claude/skills/) targets one — zero recorded loading path this cycle, tracked per §11.4.197. .opencode/ here carries only SpecKit commands (.opencode/commands/), not skills.
HelixCode SUPPORTED .claude/skills/multitrack (external tier) ADMITTED by the external-tier similarity-conflict policy gate (live-probed this run via submodules/helix_agent/cmd/similarity-probe against the real .claude/skills/, qualified identity 'Multitrack' -- not excluded on a similarity conflict). HOWEVER the FINAL registry SOURCE (live-probed this run via a real helixcode skills list) is 'builtin', NOT the external tier -- a same-named builtin copy loaded LAST overrides the admitted external-tier entry (F-22(a)); content is byte-identical post-F-13 so this is not user-visible-wrong, but claiming this skill 'resolves via the external tier' would be FALSE. helix_code/internal/agent/skill_dirs.go ExternalSkillsDir() defaults to '.claude/skills' (config-injectable via HELIXCODE_EXTERNAL_SKILLS_DIR); the external tier's LOADING MECHANISM (a single fixture skill, no similarity conflict) is REAL-BINARY verified at qa-results/hxc159/helixcode_wire/20260922T190852Z/ (green_cli_e2e.txt). HXC-159 F-17 (2026-09-23): the shared SimilarityThreshold was recalibrated 0.0256->0.0709 against the real, frontmatter-complete corpus-1 (see docs/qa/hxc159_f17_remediation_20260923T0921Z/threshold.json); the REAL 8-9-skill corpus-1 tier is nonetheless STILL policy-refused at BOTH consumers on a genuine confusable pair (action-prefix-system || reporting-workable-items, score 0.1045 -- real content overlap, not a miscalibration), captured at docs/qa/hxc159_f17_remediation_20260923T0921Z/helixcode_recalibrated_run.txt and the mirrored helix_agent_native_recalibrated_run.txt. This SUPPORTED cell's status token reports per-skill FILE PRESENCE at the platform directory (§11.4.228(A)(1)'s liveness probe), not whole-tier ACTIVATION -- as of T-P9.01 round-4 F-22(b) (2026-09-23) this note is PREFIXED, above, by a SEPARATE, live-probed verdict on the distinct fact of external-tier-SPECIFIC policy-gate admission (see external_tier_admission in catalog.json), closing the residual gap this paragraph used to describe as unclassified. Honest boundary (§11.4.6, T-P5.05.4 in specs/001-helixskills-incorporation/tasks.md): the per-skill activation-allowlist/ceiling mechanism was NOT verifiable from within helix_code/'s own file scope this cycle — grep-confirmed zero Go implementation of a per-skill activation ceiling in this module.
Gemini CLI UNTESTED N/A No .gemini/skills/ directory exists in this checkout (only .gemini/commands/, unrelated to the Agent-Skill format) and no registrar targets one — zero recorded loading path this cycle, tracked per §11.4.197 (the exact G1 gap the constitution's own §11.4.228 forensic anchor names for this platform).
Qwen Code UNTESTED N/A No .qwen/skills/ directory exists in this checkout (only .qwen/commands/ + .qwen/settings.json, unrelated to the Agent-Skill format) and no registrar targets one — zero recorded loading path this cycle, tracked per §11.4.197 (the exact G1 gap the constitution's own §11.4.228 forensic anchor names for this platform).

reporting-workable-items

  • Owner (registrar): constitution
  • Manifest source: constitution/skills/reporting-workable-items
Platform Status Loading path Evidence / Note
Claude Code SUPPORTED .claude/skills/reporting-workable-items symlink, SKILL.md read 5644 bytes (resolves to the manifest-declared source). Live-probed 2026-09-23T16:57:04Z.
OpenCode UNTESTED N/A No .opencode/skills/ directory exists in this checkout and no registrar (scripts/register_skills.sh only ever writes .claude/skills/) targets one — zero recorded loading path this cycle, tracked per §11.4.197. .opencode/ here carries only SpecKit commands (.opencode/commands/), not skills.
HelixCode SUPPORTED .claude/skills/reporting-workable-items (external tier) EXCLUDED by the external-tier similarity-conflict policy gate (live-probed this run via submodules/helix_agent/cmd/similarity-probe: qualified identity 'Reporting Workable Items' conflicts with 'Action Prefix System', score=0.1044776119402985 >= threshold=0.0709) -- this skill resolves ONLY via the SEPARATE, higher-precedence builtin tier (masking the external-tier exclusion for the end user, per F-22(a)); the prior blanket "via external tier" claim was FALSE for this skill specifically. Still SUPPORTED overall (a real tier does serve it), but NOT via the external tier. helix_code/internal/agent/skill_dirs.go ExternalSkillsDir() defaults to '.claude/skills' (config-injectable via HELIXCODE_EXTERNAL_SKILLS_DIR); the external tier's LOADING MECHANISM (a single fixture skill, no similarity conflict) is REAL-BINARY verified at qa-results/hxc159/helixcode_wire/20260922T190852Z/ (green_cli_e2e.txt). HXC-159 F-17 (2026-09-23): the shared SimilarityThreshold was recalibrated 0.0256->0.0709 against the real, frontmatter-complete corpus-1 (see docs/qa/hxc159_f17_remediation_20260923T0921Z/threshold.json); the REAL 8-9-skill corpus-1 tier is nonetheless STILL policy-refused at BOTH consumers on a genuine confusable pair (action-prefix-system || reporting-workable-items, score 0.1045 -- real content overlap, not a miscalibration), captured at docs/qa/hxc159_f17_remediation_20260923T0921Z/helixcode_recalibrated_run.txt and the mirrored helix_agent_native_recalibrated_run.txt. This SUPPORTED cell's status token reports per-skill FILE PRESENCE at the platform directory (§11.4.228(A)(1)'s liveness probe), not whole-tier ACTIVATION -- as of T-P9.01 round-4 F-22(b) (2026-09-23) this note is PREFIXED, above, by a SEPARATE, live-probed verdict on the distinct fact of external-tier-SPECIFIC policy-gate admission (see external_tier_admission in catalog.json), closing the residual gap this paragraph used to describe as unclassified. Honest boundary (§11.4.6, T-P5.05.4 in specs/001-helixskills-incorporation/tasks.md): the per-skill activation-allowlist/ceiling mechanism was NOT verifiable from within helix_code/'s own file scope this cycle — grep-confirmed zero Go implementation of a per-skill activation ceiling in this module.
Gemini CLI UNTESTED N/A No .gemini/skills/ directory exists in this checkout (only .gemini/commands/, unrelated to the Agent-Skill format) and no registrar targets one — zero recorded loading path this cycle, tracked per §11.4.197 (the exact G1 gap the constitution's own §11.4.228 forensic anchor names for this platform).
Qwen Code UNTESTED N/A No .qwen/skills/ directory exists in this checkout (only .qwen/commands/ + .qwen/settings.json, unrelated to the Agent-Skill format) and no registrar targets one — zero recorded loading path this cycle, tracked per §11.4.197 (the exact G1 gap the constitution's own §11.4.228 forensic anchor names for this platform).

scheduled-work-queue

  • Owner (registrar): constitution
  • Manifest source: constitution/skills/scheduled-work-queue
Platform Status Loading path Evidence / Note
Claude Code SUPPORTED .claude/skills/scheduled-work-queue symlink, SKILL.md read 4216 bytes (resolves to the manifest-declared source). Live-probed 2026-09-23T16:57:04Z.
OpenCode UNTESTED N/A No .opencode/skills/ directory exists in this checkout and no registrar (scripts/register_skills.sh only ever writes .claude/skills/) targets one — zero recorded loading path this cycle, tracked per §11.4.197. .opencode/ here carries only SpecKit commands (.opencode/commands/), not skills.
HelixCode SUPPORTED .claude/skills/scheduled-work-queue (external tier) ADMITTED by the external-tier similarity-conflict policy gate (live-probed this run via submodules/helix_agent/cmd/similarity-probe against the real .claude/skills/, qualified identity 'scheduled-work-queue' -- not excluded on a similarity conflict). HOWEVER the FINAL registry SOURCE (live-probed this run via a real helixcode skills list) is 'builtin', NOT the external tier -- a same-named builtin copy loaded LAST overrides the admitted external-tier entry (F-22(a)); content is byte-identical post-F-13 so this is not user-visible-wrong, but claiming this skill 'resolves via the external tier' would be FALSE. helix_code/internal/agent/skill_dirs.go ExternalSkillsDir() defaults to '.claude/skills' (config-injectable via HELIXCODE_EXTERNAL_SKILLS_DIR); the external tier's LOADING MECHANISM (a single fixture skill, no similarity conflict) is REAL-BINARY verified at qa-results/hxc159/helixcode_wire/20260922T190852Z/ (green_cli_e2e.txt). HXC-159 F-17 (2026-09-23): the shared SimilarityThreshold was recalibrated 0.0256->0.0709 against the real, frontmatter-complete corpus-1 (see docs/qa/hxc159_f17_remediation_20260923T0921Z/threshold.json); the REAL 8-9-skill corpus-1 tier is nonetheless STILL policy-refused at BOTH consumers on a genuine confusable pair (action-prefix-system || reporting-workable-items, score 0.1045 -- real content overlap, not a miscalibration), captured at docs/qa/hxc159_f17_remediation_20260923T0921Z/helixcode_recalibrated_run.txt and the mirrored helix_agent_native_recalibrated_run.txt. This SUPPORTED cell's status token reports per-skill FILE PRESENCE at the platform directory (§11.4.228(A)(1)'s liveness probe), not whole-tier ACTIVATION -- as of T-P9.01 round-4 F-22(b) (2026-09-23) this note is PREFIXED, above, by a SEPARATE, live-probed verdict on the distinct fact of external-tier-SPECIFIC policy-gate admission (see external_tier_admission in catalog.json), closing the residual gap this paragraph used to describe as unclassified. Honest boundary (§11.4.6, T-P5.05.4 in specs/001-helixskills-incorporation/tasks.md): the per-skill activation-allowlist/ceiling mechanism was NOT verifiable from within helix_code/'s own file scope this cycle — grep-confirmed zero Go implementation of a per-skill activation ceiling in this module.
Gemini CLI UNTESTED N/A No .gemini/skills/ directory exists in this checkout (only .gemini/commands/, unrelated to the Agent-Skill format) and no registrar targets one — zero recorded loading path this cycle, tracked per §11.4.197 (the exact G1 gap the constitution's own §11.4.228 forensic anchor names for this platform).
Qwen Code UNTESTED N/A No .qwen/skills/ directory exists in this checkout (only .qwen/commands/ + .qwen/settings.json, unrelated to the Agent-Skill format) and no registrar targets one — zero recorded loading path this cycle, tracked per §11.4.197 (the exact G1 gap the constitution's own §11.4.228 forensic anchor names for this platform).

session-sync

  • Owner (registrar): constitution
  • Manifest source: constitution/skills/session-sync
Platform Status Loading path Evidence / Note
Claude Code SUPPORTED .claude/skills/session-sync symlink, SKILL.md read 5473 bytes (resolves to the manifest-declared source). Live-probed 2026-09-23T16:57:04Z.
OpenCode UNTESTED N/A No .opencode/skills/ directory exists in this checkout and no registrar (scripts/register_skills.sh only ever writes .claude/skills/) targets one — zero recorded loading path this cycle, tracked per §11.4.197. .opencode/ here carries only SpecKit commands (.opencode/commands/), not skills.
HelixCode SUPPORTED .claude/skills/session-sync (external tier) EXCLUDED by the external-tier similarity-conflict policy gate (live-probed this run via submodules/helix_agent/cmd/similarity-probe: qualified identity 'session-sync' conflicts with 'Multitrack', score=0.07936507936507936 >= threshold=0.0709) -- this skill resolves ONLY via the SEPARATE, higher-precedence builtin tier (masking the external-tier exclusion for the end user, per F-22(a)); the prior blanket "via external tier" claim was FALSE for this skill specifically. Still SUPPORTED overall (a real tier does serve it), but NOT via the external tier. helix_code/internal/agent/skill_dirs.go ExternalSkillsDir() defaults to '.claude/skills' (config-injectable via HELIXCODE_EXTERNAL_SKILLS_DIR); the external tier's LOADING MECHANISM (a single fixture skill, no similarity conflict) is REAL-BINARY verified at qa-results/hxc159/helixcode_wire/20260922T190852Z/ (green_cli_e2e.txt). HXC-159 F-17 (2026-09-23): the shared SimilarityThreshold was recalibrated 0.0256->0.0709 against the real, frontmatter-complete corpus-1 (see docs/qa/hxc159_f17_remediation_20260923T0921Z/threshold.json); the REAL 8-9-skill corpus-1 tier is nonetheless STILL policy-refused at BOTH consumers on a genuine confusable pair (action-prefix-system || reporting-workable-items, score 0.1045 -- real content overlap, not a miscalibration), captured at docs/qa/hxc159_f17_remediation_20260923T0921Z/helixcode_recalibrated_run.txt and the mirrored helix_agent_native_recalibrated_run.txt. This SUPPORTED cell's status token reports per-skill FILE PRESENCE at the platform directory (§11.4.228(A)(1)'s liveness probe), not whole-tier ACTIVATION -- as of T-P9.01 round-4 F-22(b) (2026-09-23) this note is PREFIXED, above, by a SEPARATE, live-probed verdict on the distinct fact of external-tier-SPECIFIC policy-gate admission (see external_tier_admission in catalog.json), closing the residual gap this paragraph used to describe as unclassified. Honest boundary (§11.4.6, T-P5.05.4 in specs/001-helixskills-incorporation/tasks.md): the per-skill activation-allowlist/ceiling mechanism was NOT verifiable from within helix_code/'s own file scope this cycle — grep-confirmed zero Go implementation of a per-skill activation ceiling in this module.
Gemini CLI UNTESTED N/A No .gemini/skills/ directory exists in this checkout (only .gemini/commands/, unrelated to the Agent-Skill format) and no registrar targets one — zero recorded loading path this cycle, tracked per §11.4.197 (the exact G1 gap the constitution's own §11.4.228 forensic anchor names for this platform).
Qwen Code UNTESTED N/A No .qwen/skills/ directory exists in this checkout (only .qwen/commands/ + .qwen/settings.json, unrelated to the Agent-Skill format) and no registrar targets one — zero recorded loading path this cycle, tracked per §11.4.197 (the exact G1 gap the constitution's own §11.4.228 forensic anchor names for this platform).

skill-catalog

  • Owner (registrar): constitution
  • Manifest source: constitution/skills/skill-catalog
Platform Status Loading path Evidence / Note
Claude Code SUPPORTED .claude/skills/skill-catalog symlink, SKILL.md read 2328 bytes (resolves to the manifest-declared source). Live-probed 2026-09-23T16:57:04Z.
OpenCode UNTESTED N/A No .opencode/skills/ directory exists in this checkout and no registrar (scripts/register_skills.sh only ever writes .claude/skills/) targets one — zero recorded loading path this cycle, tracked per §11.4.197. .opencode/ here carries only SpecKit commands (.opencode/commands/), not skills.
HelixCode SUPPORTED .claude/skills/skill-catalog (external tier) ADMITTED by the external-tier similarity-conflict policy gate (live-probed this run via submodules/helix_agent/cmd/similarity-probe against the real .claude/skills/, qualified identity 'skill-catalog' -- not excluded on a similarity conflict). The FINAL registry SOURCE (live-probed this run via a real helixcode skills list) CONFIRMS it -- genuinely served from the external tier -> .claude/skills/skill-catalog. helix_code/internal/agent/skill_dirs.go ExternalSkillsDir() defaults to '.claude/skills' (config-injectable via HELIXCODE_EXTERNAL_SKILLS_DIR); the external tier's LOADING MECHANISM (a single fixture skill, no similarity conflict) is REAL-BINARY verified at qa-results/hxc159/helixcode_wire/20260922T190852Z/ (green_cli_e2e.txt). HXC-159 F-17 (2026-09-23): the shared SimilarityThreshold was recalibrated 0.0256->0.0709 against the real, frontmatter-complete corpus-1 (see docs/qa/hxc159_f17_remediation_20260923T0921Z/threshold.json); the REAL 8-9-skill corpus-1 tier is nonetheless STILL policy-refused at BOTH consumers on a genuine confusable pair (action-prefix-system || reporting-workable-items, score 0.1045 -- real content overlap, not a miscalibration), captured at docs/qa/hxc159_f17_remediation_20260923T0921Z/helixcode_recalibrated_run.txt and the mirrored helix_agent_native_recalibrated_run.txt. This SUPPORTED cell's status token reports per-skill FILE PRESENCE at the platform directory (§11.4.228(A)(1)'s liveness probe), not whole-tier ACTIVATION -- as of T-P9.01 round-4 F-22(b) (2026-09-23) this note is PREFIXED, above, by a SEPARATE, live-probed verdict on the distinct fact of external-tier-SPECIFIC policy-gate admission (see external_tier_admission in catalog.json), closing the residual gap this paragraph used to describe as unclassified. Honest boundary (§11.4.6, T-P5.05.4 in specs/001-helixskills-incorporation/tasks.md): the per-skill activation-allowlist/ceiling mechanism was NOT verifiable from within helix_code/'s own file scope this cycle — grep-confirmed zero Go implementation of a per-skill activation ceiling in this module.
Gemini CLI UNTESTED N/A No .gemini/skills/ directory exists in this checkout (only .gemini/commands/, unrelated to the Agent-Skill format) and no registrar targets one — zero recorded loading path this cycle, tracked per §11.4.197 (the exact G1 gap the constitution's own §11.4.228 forensic anchor names for this platform).
Qwen Code UNTESTED N/A No .qwen/skills/ directory exists in this checkout (only .qwen/commands/ + .qwen/settings.json, unrelated to the Agent-Skill format) and no registrar targets one — zero recorded loading path this cycle, tracked per §11.4.197 (the exact G1 gap the constitution's own §11.4.228 forensic anchor names for this platform).

workable-item-lifecycle

  • Owner (registrar): constitution
  • Manifest source: constitution/skills/workable-item-lifecycle
Platform Status Loading path Evidence / Note
Claude Code SUPPORTED .claude/skills/workable-item-lifecycle symlink, SKILL.md read 5297 bytes (resolves to the manifest-declared source). Live-probed 2026-09-23T16:57:04Z.
OpenCode UNTESTED N/A No .opencode/skills/ directory exists in this checkout and no registrar (scripts/register_skills.sh only ever writes .claude/skills/) targets one — zero recorded loading path this cycle, tracked per §11.4.197. .opencode/ here carries only SpecKit commands (.opencode/commands/), not skills.
HelixCode SUPPORTED .claude/skills/workable-item-lifecycle (external tier) ADMITTED by the external-tier similarity-conflict policy gate (live-probed this run via submodules/helix_agent/cmd/similarity-probe against the real .claude/skills/, qualified identity 'Workable Item Lifecycle' -- not excluded on a similarity conflict). HOWEVER the FINAL registry SOURCE (live-probed this run via a real helixcode skills list) is 'builtin', NOT the external tier -- a same-named builtin copy loaded LAST overrides the admitted external-tier entry (F-22(a)); content is byte-identical post-F-13 so this is not user-visible-wrong, but claiming this skill 'resolves via the external tier' would be FALSE. helix_code/internal/agent/skill_dirs.go ExternalSkillsDir() defaults to '.claude/skills' (config-injectable via HELIXCODE_EXTERNAL_SKILLS_DIR); the external tier's LOADING MECHANISM (a single fixture skill, no similarity conflict) is REAL-BINARY verified at qa-results/hxc159/helixcode_wire/20260922T190852Z/ (green_cli_e2e.txt). HXC-159 F-17 (2026-09-23): the shared SimilarityThreshold was recalibrated 0.0256->0.0709 against the real, frontmatter-complete corpus-1 (see docs/qa/hxc159_f17_remediation_20260923T0921Z/threshold.json); the REAL 8-9-skill corpus-1 tier is nonetheless STILL policy-refused at BOTH consumers on a genuine confusable pair (action-prefix-system || reporting-workable-items, score 0.1045 -- real content overlap, not a miscalibration), captured at docs/qa/hxc159_f17_remediation_20260923T0921Z/helixcode_recalibrated_run.txt and the mirrored helix_agent_native_recalibrated_run.txt. This SUPPORTED cell's status token reports per-skill FILE PRESENCE at the platform directory (§11.4.228(A)(1)'s liveness probe), not whole-tier ACTIVATION -- as of T-P9.01 round-4 F-22(b) (2026-09-23) this note is PREFIXED, above, by a SEPARATE, live-probed verdict on the distinct fact of external-tier-SPECIFIC policy-gate admission (see external_tier_admission in catalog.json), closing the residual gap this paragraph used to describe as unclassified. Honest boundary (§11.4.6, T-P5.05.4 in specs/001-helixskills-incorporation/tasks.md): the per-skill activation-allowlist/ceiling mechanism was NOT verifiable from within helix_code/'s own file scope this cycle — grep-confirmed zero Go implementation of a per-skill activation ceiling in this module.
Gemini CLI UNTESTED N/A No .gemini/skills/ directory exists in this checkout (only .gemini/commands/, unrelated to the Agent-Skill format) and no registrar targets one — zero recorded loading path this cycle, tracked per §11.4.197 (the exact G1 gap the constitution's own §11.4.228 forensic anchor names for this platform).
Qwen Code UNTESTED N/A No .qwen/skills/ directory exists in this checkout (only .qwen/commands/ + .qwen/settings.json, unrelated to the Agent-Skill format) and no registrar targets one — zero recorded loading path this cycle, tracked per §11.4.197 (the exact G1 gap the constitution's own §11.4.228 forensic anchor names for this platform).

Source tracking

Per §11.4.228(B): every extension's canonical source is machine-readable via scripts/skills_manifest.tsv itself (name, repo-relative source directory, owning registrar) — the manifest row IS this catalogue's EXTENSION_SOURCE record for every entry here (no separate EXTENSION_SOURCE.yaml per skill; the manifest already carries the same three fields at repo scope, avoiding a duplicated, driftable second record per §11.4.106). All eight entries are constitution-owned (owner=constitution) and SYMLINKED (never copied) from constitution/skills/<name>/ into .claude/skills/<name> by scripts/register_skills.sh — the preferred-over-copies provenance §11.4.228(B)(2) requires.

Source manifest: /home/milosvasic/Projects/helix_code/scripts/skills_manifest.tsv (8 rows).