AgencyAI

Snapshot Jul 25, 2026codex/agencyai-harness-audit · 8f66e21Source cleanBundle sync passRead-only snapshot

Snapshot proposals

Audit findings

Expand a row for evidence and a copyable fix request. Nothing runs automatically.

Deduplicated proposal range2.2K–3.8K wordsDeduplicated maximum proposal range per skill. Estimates can overlap and are not additive.

mediumtcm-client-site-startup is close to the skill size limitCTX-LINE-CEILING-TCM-CLIENT-SITE-STARTUPProgressive disclosure

What this means

This skill has 492 lines and 5,310 words. The current check watches line count, but not how dense those lines are.

Why it matters

A word-heavy skill can pass the 500-line check while still loading too much context.

Recommended next step

Add word and token checks, then move detailed sections behind explicit reference links after approval.

Nothing will change automatically.This finding remains a proposal until its stable ID is explicitly approved.

Copies only. Paste the request into Codex to approve and run it.

Technical details

Finding ID: CTX-LINE-CEILING-TCM-CLIENT-SITE-STARTUP
Category: Progressive disclosure
Approval: pending
Original title: tcm-client-site-startup sits near the structural line ceiling
Technical summary: The existing 500-line guard can pass while a skill still contributes a large instruction payload.
Recommendation: Add word/token and routed-reference checks rather than relying on line count alone.
Evidence: 492 lines
5310 words

mediumtcm-ghl-form-tracking is close to the skill size limitCTX-LINE-CEILING-TCM-GHL-FORM-TRACKINGProgressive disclosure

What this means

This skill has 489 lines and 4,959 words. The current check watches line count, but not how dense those lines are.

Why it matters

A word-heavy skill can pass the 500-line check while still loading too much context.

Recommended next step

Add word and token checks, then move detailed sections behind explicit reference links after approval.

Nothing will change automatically.This finding remains a proposal until its stable ID is explicitly approved.

Copies only. Paste the request into Codex to approve and run it.

Technical details

Finding ID: CTX-LINE-CEILING-TCM-GHL-FORM-TRACKING
Category: Progressive disclosure
Approval: pending
Original title: tcm-ghl-form-tracking sits near the structural line ceiling
Technical summary: The existing 500-line guard can pass while a skill still contributes a large instruction payload.
Recommendation: Add word/token and routed-reference checks rather than relying on line count alone.
Evidence: 489 lines
4959 words

mediumtcm-visual-quality is close to the skill size limitCTX-LINE-CEILING-TCM-VISUAL-QUALITYProgressive disclosure

What this means

This skill has 480 lines and 3,498 words. The current check watches line count, but not how dense those lines are.

Why it matters

A word-heavy skill can pass the 500-line check while still loading too much context.

Recommended next step

Add word and token checks, then move detailed sections behind explicit reference links after approval.

Nothing will change automatically.This finding remains a proposal until its stable ID is explicitly approved.

Copies only. Paste the request into Codex to approve and run it.

Technical details

Finding ID: CTX-LINE-CEILING-TCM-VISUAL-QUALITY
Category: Progressive disclosure
Approval: pending
Original title: tcm-visual-quality sits near the structural line ceiling
Technical summary: The existing 500-line guard can pass while a skill still contributes a large instruction payload.
Recommendation: Add word/token and routed-reference checks rather than relying on line count alone.
Evidence: 480 lines
3498 words

mediumhubspot-crm and salesforce-crm repeat some of the same instructionsCTX-OVERLAP-HUBSPOT-CRM-SALESFORCE-CRMSkill-body similarity

What this means

These two skills contain many of the same instruction phrases. The overlap measurement is 19.7% for the smaller skill.

Why it matters

Repeated rules increase context use and can drift into conflicting versions over time.

Recommended next step

After approval, choose one skill as the owner of each shared rule and have the other route to it.

Possible context reduction0–17 wordsThis is an estimate for this finding, not a guaranteed or additive saving.

Nothing will change automatically.This finding remains a proposal until its stable ID is explicitly approved.

Copies only. Paste the request into Codex to approve and run it.

Technical details

Finding ID: CTX-OVERLAP-HUBSPOT-CRM-SALESFORCE-CRM
Category: Skill-body similarity
Approval: pending
Original title: hubspot-crm overlaps salesforce-crm
Technical summary: 19.7% of the smaller skill's five-word sequences also appear in the paired skill.
Recommendation: Assign one authoritative owner for shared rules and replace repeated detail with explicit routing.
Evidence: 30 shared five-word sequences
This measures body similarity, not trigger collision
Savings scope: pair:hubspot-crm|salesforce-crm

mediumtcm-client-site-startup and tcm-visual-quality repeat some of the same instructionsCTX-OVERLAP-TCM-CLIENT-SITE-STARTUP-TCM-VISUAL-QUALITYSkill-body similarity

What this means

These two skills contain many of the same instruction phrases. The overlap measurement is 16.7% for the smaller skill.

Why it matters

Repeated rules increase context use and can drift into conflicting versions over time.

Recommended next step

After approval, choose one skill as the owner of each shared rule and have the other route to it.

Possible context reduction0–44 wordsThis is an estimate for this finding, not a guaranteed or additive saving.

Nothing will change automatically.This finding remains a proposal until its stable ID is explicitly approved.

Copies only. Paste the request into Codex to approve and run it.

Technical details

Finding ID: CTX-OVERLAP-TCM-CLIENT-SITE-STARTUP-TCM-VISUAL-QUALITY
Category: Skill-body similarity
Approval: pending
Original title: tcm-client-site-startup overlaps tcm-visual-quality
Technical summary: 16.7% of the smaller skill's five-word sequences also appear in the paired skill.
Recommendation: Assign one authoritative owner for shared rules and replace repeated detail with explicit routing.
Evidence: 80 shared five-word sequences
This measures body similarity, not trigger collision
Savings scope: pair:tcm-client-site-startup|tcm-visual-quality

mediumtcm-client-site-startup loads a large set of instructions at onceCTX-OVERSIZE-TCM-CLIENT-SITE-STARTUPContext cost

What this means

Selecting tcm-client-site-startup loads 5,310 words before Codex opens any supporting references.

Why it matters

That uses a large share of the working context before the task itself begins, which can make important requirements easier to miss.

Recommended next step

After approval, keep the essential operating rules in the skill and move phase-specific detail into references that load only when needed.

Possible context reduction983–1,686 wordsThis is an estimate for this finding, not a guaranteed or additive saving.

Nothing will change automatically.This finding remains a proposal until its stable ID is explicitly approved.

Copies only. Paste the request into Codex to approve and run it.

Technical details

Finding ID: CTX-OVERSIZE-TCM-CLIENT-SITE-STARTUP
Category: Context cost
Approval: pending
Original title: tcm-client-site-startup carries a large triggered payload
Technical summary: 5,310 words load when the skill is selected, before any referenced detail is opened.
Recommendation: Move phase-specific detail behind explicit reference routing while keeping the top-level operating contract decision-complete.
Evidence: skills/tcm-client-site-startup/SKILL.md: 5310 words, 492 lines
1 linked local references
Savings scope: skill:tcm-client-site-startup

mediumtcm-ghl-form-tracking loads a large set of instructions at onceCTX-OVERSIZE-TCM-GHL-FORM-TRACKINGContext cost

What this means

Selecting tcm-ghl-form-tracking loads 4,959 words before Codex opens any supporting references.

Why it matters

That uses a large share of the working context before the task itself begins, which can make important requirements easier to miss.

Recommended next step

After approval, keep the essential operating rules in the skill and move phase-specific detail into references that load only when needed.

Possible context reduction861–1,475 wordsThis is an estimate for this finding, not a guaranteed or additive saving.

Nothing will change automatically.This finding remains a proposal until its stable ID is explicitly approved.

Copies only. Paste the request into Codex to approve and run it.

Technical details

Finding ID: CTX-OVERSIZE-TCM-GHL-FORM-TRACKING
Category: Context cost
Approval: pending
Original title: tcm-ghl-form-tracking carries a large triggered payload
Technical summary: 4,959 words load when the skill is selected, before any referenced detail is opened.
Recommendation: Move phase-specific detail behind explicit reference routing while keeping the top-level operating contract decision-complete.
Evidence: skills/tcm-ghl-form-tracking/SKILL.md: 4959 words, 489 lines
2 linked local references
Savings scope: skill:tcm-ghl-form-tracking

mediumtcm-visual-quality loads a large set of instructions at onceCTX-OVERSIZE-TCM-VISUAL-QUALITYContext cost

What this means

Selecting tcm-visual-quality loads 3,498 words before Codex opens any supporting references.

Why it matters

That uses a large share of the working context before the task itself begins, which can make important requirements easier to miss.

Recommended next step

After approval, keep the essential operating rules in the skill and move phase-specific detail into references that load only when needed.

Possible context reduction349–599 wordsThis is an estimate for this finding, not a guaranteed or additive saving.

Nothing will change automatically.This finding remains a proposal until its stable ID is explicitly approved.

Copies only. Paste the request into Codex to approve and run it.

Technical details

Finding ID: CTX-OVERSIZE-TCM-VISUAL-QUALITY
Category: Context cost
Approval: pending
Original title: tcm-visual-quality carries a large triggered payload
Technical summary: 3,498 words load when the skill is selected, before any referenced detail is opened.
Recommendation: Move phase-specific detail behind explicit reference routing while keeping the top-level operating contract decision-complete.
Evidence: skills/tcm-visual-quality/SKILL.md: 3498 words, 480 lines
1 linked local references
Savings scope: skill:tcm-visual-quality

mediumMemory points to an AgencyAI audit skill that no longer existsMEM-STALE-REFERENCE-001Memory hygiene

What this means

The memory registry still names an older AgencyAI audit skill, but that skill is no longer installed.

Why it matters

A future audit request could be routed toward a workflow that Codex cannot load.

Recommended next step

After approval, submit an ad-hoc memory correction note that removes or replaces the stale pointer. Do not edit memory files directly.

Nothing will change automatically.This finding remains a proposal until its stable ID is explicitly approved.

Copies only. Paste the request into Codex to approve and run it.

Technical details

Finding ID: MEM-STALE-REFERENCE-001
Category: Memory hygiene
Approval: pending
Original title: Memory registry points to a removed AgencyAI audit skill
Technical summary: The memory registry retains a related-skill pointer that no longer resolves in AgencyAI.
Recommendation: If approved, submit an ad-hoc memory correction note; never edit memory files directly.
Evidence: Memory contains the removed relative skill path
No matching skill directory exists in the current plugin

mediumtcm-client-site-startup and tcm-domain-rebuild-deploy may activate for the same requestTRIGGER-COLLISION-TCM-CLIENT-SITE-STARTUP-TCM-DOMAIN-REBUILD-DEPLOYTrigger collision

What this means

Their skill descriptions use enough of the same trigger terms that Codex may consider both workflows for one task.

Why it matters

Codex may choose the wrong workflow or load more instructions than needed.

Recommended next step

Give each skill clearer, mutually exclusive trigger wording or add an explicit router after approval.

Nothing will change automatically.This finding remains a proposal until its stable ID is explicitly approved.

Copies only. Paste the request into Codex to approve and run it.

Technical details

Finding ID: TRIGGER-COLLISION-TCM-CLIENT-SITE-STARTUP-TCM-DOMAIN-REBUILD-DEPLOY
Category: Trigger collision
Approval: pending
Original title: tcm-client-site-startup and tcm-domain-rebuild-deploy have overlapping triggers
Technical summary: 39.3% trigger-term containment across the pair's skill descriptions.
Recommendation: Clarify mutually exclusive trigger boundaries or designate an explicit router.
Evidence: Shared trigger terms: client, domain, generation, industry, lead, ppc, research, service, site, source, website
23.4% Jaccard similarity

mediumtcm-reference-screenshot-archive and tcm-visual-screenshot-review may activate for the same requestTRIGGER-COLLISION-TCM-REFERENCE-SCREENSHOT-ARCHIVE-TCM-VISUAL-SCREENSHOT-REVIEWTrigger collision

What this means

Their skill descriptions use enough of the same trigger terms that Codex may consider both workflows for one task.

Why it matters

Codex may choose the wrong workflow or load more instructions than needed.

Recommended next step

Give each skill clearer, mutually exclusive trigger wording or add an explicit router after approval.

Nothing will change automatically.This finding remains a proposal until its stable ID is explicitly approved.

Copies only. Paste the request into Codex to approve and run it.

Technical details

Finding ID: TRIGGER-COLLISION-TCM-REFERENCE-SCREENSHOT-ARCHIVE-TCM-VISUAL-SCREENSHOT-REVIEW
Category: Trigger collision
Approval: pending
Original title: tcm-reference-screenshot-archive and tcm-visual-screenshot-review have overlapping triggers
Technical summary: 47.8% trigger-term containment across the pair's skill descriptions.
Recommendation: Clarify mutually exclusive trigger boundaries or designate an explicit router.
Evidence: Shared trigger terms: archives, back, datasets, for, page, playwright, reference, roofing, screenshot, screenshots, site
28.2% Jaccard similarity

mediumCurrent validation does not check instruction sizeVALIDATION-CONTEXT-COVERAGE-001Deterministic checks

What this means

The structural audit checks files, links, duplicate sections, and line counts, but it does not fail on word count, estimated tokens, or instruction overlap.

Why it matters

A skill can pass validation even when it loads too much context or repeats another skill.

Recommended next step

Keep these findings advisory until thresholds are approved, then add the approved limits to CI.

Nothing will change automatically.This finding remains a proposal until its stable ID is explicitly approved.

Copies only. Paste the request into Codex to approve and run it.

Technical details

Finding ID: VALIDATION-CONTEXT-COVERAGE-001
Category: Deterministic checks
Approval: pending
Original title: Structural validation does not enforce context-cost limits
Technical summary: The existing audit checks packaging, links, exact duplicate sections, and line ceilings but not words, estimated tokens, or semantic overlap.
Recommendation: Keep this harness audit as a separate approval surface, then promote approved thresholds into CI.
Evidence: tools/audit-plugin-structure.mjs has no context-cost metric
Current structural audit can pass oversized skills

Workflows

Workflow loading

What each workflow does and the prompt context it loads.

Turns a business domain into a complete client onboarding package, including the website, PPC pages, tracking, and deployment.

Example promptOnboard examplebusiness.com

Estimated visible words28,115≈ 37.4K estimated tokens

Model checkPASS · 6/6 assertionsPrimarytcm-domain-rebuild-deployCompanionscodex-working-contract · tcm-client-site-startup · tcm-full-client-site · roofing-ppc-landers · tcm-visual-quality · tcm-ghl-form-tracking · tcm-page-speed

✓ Routes primarily to tcm-domain-rebuild-deploy✓ Includes tcm-domain-rebuild-deploy✓ Includes tcm-client-site-startup✓ Includes tcm-full-client-site✓ Includes roofing-ppc-landers✓ Includes tcm-ghl-form-tracking

Measured active context only. Estimates exclude opaque platform instructions, searched memory, task-opened references, and invoked tool source.

Available skill payloads

Skill context footprint

Highlighted skills are active for One-line full onboarding.

tcm-client-site-startupSite delivery

5,3101 refs

tcm-ghl-form-trackingTracking and CRM

4,9592 refs

tcm-visual-qualityVisual and quality

3,4981 refs

tcm-full-client-siteSite delivery

2,7960 refs

tcm-meta-ads-mockupsAds and policy

2,6270 refs

tcm-domain-rebuild-deploySite delivery

1,8383 refs

tcm-page-speedVisual and quality

1,7700 refs

tcm-client-removalOperations

1,2800 refs

tcm-ghl-ui-assetsTracking and CRM

1,2720 refs

roofing-ppc-landersSite delivery

1,2293 refs

tcm-meta-ads-mcpAds and policy

1,2261 refs

tcm-authenticated-chromeOther

1,0890 refs

sites-managed-gitSite delivery

8890 refs

codex-working-contractOperations

7920 refs

tcm-visual-screenshot-reviewVisual and quality

7660 refs

tcm-ga4-client-analyticsTracking and CRM

7340 refs

tcm-reference-screenshot-archiveVisual and quality

5790 refs

hubspot-crmExternal services

5020 refs

tracking-and-form-safetyTracking and CRM

4800 refs

tcm-onboarding-learning-loopOperations

3850 refs

ahrefsExternal services

3480 refs

salesforce-crmExternal services

3300 refs

chatgpt-handoffOperations

2390 refs

On-demand context

Largest reference surfaces

execution-details.mdOn demand

8,096

TCM_ONBOARDING_JOB_TEMPLATE.jsonOn demand

5,782

startup-defaults.mdOn demand

2,622

rate-limits-and-access-tiers.mdOn demand

2,021

TCM_BUILD_BRIEF_TEMPLATE.mdOn demand

1,716

ONBOARDING_COORDINATOR_AGENT_PROMPT.mdOn demand

1,676

ONBOARDING_ADS_BLUEPRINT_AGENT_PROMPT.mdOn demand

1,520

representative-dental-2026-07-12.mdOn demand

1,412

validation-and-visual-details.mdOn demand

1,391

offer-design-and-copy-details.mdOn demand

1,336

coordination-and-gates.mdOn demand

1,323

api-usage-rules.mdOn demand

1,199

Separated evidence

Overlap evidence

Trigger-collision matrix

2 flagged trigger collisions; cells show containment.
Skill	1	2	3	4
1	—	39.3%	·	·
2	39.3%	—	·	·
3	·	·	—	47.8%
4	·	·	47.8%	—

tcm-client-site-startup
tcm-domain-rebuild-deploy
tcm-reference-screenshot-archive
tcm-visual-screenshot-review

Skill-body five-gram similarity

Instruction-body similarity is not a trigger collision and is shown separately.

hubspot-crm↔salesforce-crm

19.7%30 shared 5-grams · smaller salesforce-crm

tcm-client-site-startup↔tcm-visual-quality

16.7%80 shared 5-grams · smaller tcm-visual-quality

ahrefs↔salesforce-crm

11.8%18 shared 5-grams · smaller salesforce-crm

tcm-client-site-startup↔tcm-ghl-form-tracking

9.2%80 shared 5-grams · smaller tcm-client-site-startup

ahrefs↔hubspot-crm

8.5%17 shared 5-grams · smaller ahrefs

roofing-ppc-landers↔tcm-visual-quality

5.6%27 shared 5-grams · smaller tcm-visual-quality

tcm-ghl-form-tracking↔tcm-visual-quality

4.2%20 shared 5-grams · smaller tcm-visual-quality

tcm-authenticated-chrome↔tcm-full-client-site

4.1%22 shared 5-grams · smaller tcm-authenticated-chrome

roofing-ppc-landers↔tcm-full-client-site

3.9%28 shared 5-grams · smaller roofing-ppc-landers

roofing-ppc-landers↔tcm-ghl-form-tracking

3.4%24 shared 5-grams · smaller roofing-ppc-landers

tcm-authenticated-chrome↔tcm-ghl-ui-assets

3.1%13 shared 5-grams · smaller tcm-ghl-ui-assets

tcm-domain-rebuild-deploy↔tcm-visual-quality

2.9%14 shared 5-grams · smaller tcm-visual-quality

Remediation history

Approved remediation receipts

Historical proof of approved changes, scope isolation, and validation results. These receipts do not reduce the current finding count.

6 historical receipts

ResolvedAn old skill name is still in your instructionsCFG-STALE-SKILL-0011716findingsJul 15, 2026

Replaced the retired global GHL skill name with tcm-ghl-form-tracking while preserving the existing deprecation guard.

Before17 findings3 high priority

After16 findings2 high priority

Only CFG-STALE-SKILL-001 was remediated. No other audit finding was applied.

Validation receipt

Pass
AgencyAI checksPlugin validation, audit tests, structural audit, and secret scan passed.
Pass
Routing probesAll 5 modeled task scenarios passed.
Pass
Source and runtime parity203 files compared with 0 missing and 0 different.
Pass
Deprecation guardThe guard remains unchanged and continues to block the retired identifier.

Resolvedtcm-client-site-startup and tcm-domain-rebuild-deploy repeat some of the same instructionsCTX-OVERLAP-TCM-CLIENT-SITE-STARTUP-TCM-DOMAIN-REBUILD-DEPLOY1615findingsJul 15, 2026

Made tcm-client-site-startup authoritative for source facts and build-brief inputs, made tcm-domain-rebuild-deploy authoritative for onboarding execution, and routed the startup handoff to the orchestrator.

Before16 findings2 high priority

After15 findings1 high priority

Only CTX-OVERLAP-TCM-CLIENT-SITE-STARTUP-TCM-DOMAIN-REBUILD-DEPLOY was remediated. No other audit finding was applied.

Validation receipt

Pass
AgencyAI checksPlugin validation, audit tests, structural audit, and secret scan passed.
Pass
Routing probesAll 5 modeled task scenarios passed.
Pass
Source and runtime parity203 files compared with 0 missing and 0 different.
Pass
Finding isolationThe finding-set comparison removed only the approved stable ID and added no findings.

Resolvedtcm-domain-rebuild-deploy loads a large set of instructions at onceCTX-OVERSIZE-TCM-DOMAIN-REBUILD-DEPLOY1514findingsJul 15, 2026

Reduced the triggered tcm-domain-rebuild-deploy skill from 7,844 to 1,781 words by keeping essential operating decisions in the skill and routing phase-specific coordination and execution detail to on-demand references.

Before15 findings1 high priority

After14 findings0 high priority

Only CTX-OVERSIZE-TCM-DOMAIN-REBUILD-DEPLOY was remediated. No other audit finding was applied.

Validation receipt

Pass
Triggered contextThe selected skill now loads 1,781 words instead of 7,844, a reduction of 6,063 words.
Pass
AgencyAI checksPlugin validation, audit tests, structural audit, and secret scan passed.
Pass
Routing probesAll 5 modeled task scenarios passed after the progressive-disclosure refactor.
Pass
Source and runtime parity207 files were compared with 0 missing and 0 different.
Pass
Finding isolationThe finding-set comparison removed only the approved stable ID and added no findings.

ResolvedYour instructions point to an old shared-workflow folderCFG-STALE-PATH-0011413findingsJul 15, 2026

Replaced the old shared-workflow folder in global guidance with AgencyAI as the canonical editable source, while explicitly keeping live applications, client sites, and deployment runtimes in their separate repositories.

Before14 findings0 high priority

After13 findings0 high priority

Only CFG-STALE-PATH-001 was remediated. No other audit finding was applied.

Validation receipt

Pass
Authority guidanceGlobal guidance now names AgencyAI as the canonical shared plugin source and keeps runtime repositories separate.
Pass
AgencyAI checksPlugin validation, audit tests, structural audit, and secret scan passed.
Pass
Routing probesAll 5 modeled task scenarios passed after the authority correction.
Pass
Source and runtime parity207 files were compared with 0 missing and 0 different.
Pass
Finding isolationThe finding-set comparison removed only the approved stable ID and added no findings.

Resolvedroofing-ppc-landers is close to the skill size limitCTX-LINE-CEILING-ROOFING-PPC-LANDERS1312findingsJul 15, 2026

Reduced the triggered roofing-ppc-landers skill from 485 lines and 2,994 audited words to 150 lines and 1,269 audited words, with offer, design, social-proof, copy, and form implementation detail routed to an explicit on-demand reference.

Before13 findings0 high priority

After12 findings0 high priority

Only CTX-LINE-CEILING-ROOFING-PPC-LANDERS was remediated. No other audit finding was applied.

Validation receipt

Pass
Skill context ceilingA deterministic regression test now enforces at most 2,499 words and 3,324 estimated tokens; the skill measures 1,269 words and 1,688 estimated tokens.
Pass
Routed referencesThe main skill directly links the offer/design/copy details, validation/visual details, and canonical PPC form/copy contract used by later phases.
Pass
AgencyAI checksPlugin validation, audit tests, structural audit, skill validation, and secret scan passed.
Pass
Routing probesAll 5 modeled task scenarios passed after the progressive-disclosure refactor.
Pass
Source and runtime parity209 files were compared with 0 missing and 0 different.
Pass
Finding isolationThe finding-set comparison removed only the approved stable ID, added no findings, and retained the separate validation-coverage finding.

Resolvedtcm-meta-ads-mcp and tcm-meta-ads-mockups may activate for the same requestTRIGGER-COLLISION-TCM-META-ADS-MCP-TCM-META-ADS-MOCKUPS1312findingsJul 17, 2026

Separated the activation contract by lifecycle phase: tcm-meta-ads-mcp owns new-object creation plus verification of that batch's returned IDs, while tcm-meta-ads-mockups owns blueprint governance, standalone readback and audit, creation-surface selection, and approved removal.

Before13 findings0 high priority

After12 findings0 high priority

Only TRIGGER-COLLISION-TCM-META-ADS-MCP-TCM-META-ADS-MOCKUPS was remediated. No other audit finding was applied.

Validation receipt

Pass
Mutually exclusive routingThe two frontmatter descriptions and skill bodies now prohibit parallel activation and define separate creation, governance, standalone verification, and removal phases.
Pass
Trigger collision metricContainment fell from 37.5% to 23.8% and Jaccard similarity fell from 21.4% to 13.5%; the audited pair now reports collision=false.
Pass
Routing probesAll 7 modeled scenarios passed; Meta governance and lifecycle probes exclude the MCP skill, while new-object creation excludes the mockups skill.
Pass
AgencyAI checksPlugin validation, full audit tests, structural audit, both edited-skill validators, and the secret scan passed.
Pass
Source and runtime parity213 files were compared with 0 untracked source files, 0 missing cache files, 0 different files, 0 cache-only files, and 0 stale skill junctions.
Pass
Finding isolationThe finding-set comparison removed only the approved stable ID and added no findings.

AgencyAI

How AgencyAI assembles context

Instruction authority

Session & routing

Conditional context

Invoked resources

Indexed context footprint

Measurement

Loading classes

Audit findings

Workflow loading

Skill context footprint

Largest reference surfaces

Overlap evidence

Trigger-collision matrix

Skill-body five-gram similarity

Approved remediation receipts