MCP Governance Audit

{% if audit %}

{{ audit.gated_current }} / {{ audit.current_count }}

Tools Gated

Audit Status

Mutators Ungated (ratchet target)

Last Run

{% if has_drift %}

Drift findings — manual review

REGRESSION: tools lost their gate: {{ audit.findings.gate_drop_outs | join(', ') }}
NEW: unclassified tools: {{ audit.findings.new_unclassified_tools | join(', ') }}
RATCHET: mutators that gained a gate (move to gated in baseline): {{ audit.findings.gate_added_ratchet_candidates | join(', ') }}
REMOVED tools no longer present (T-1166 deprecation may have completed): {{ audit.findings.removed_tools | join(', ') }}

TAG-FORMAT DRIFT (T-1649): live sessions carry non-canonical tag prefixes — orchestrator routing silently falls back to defaults. T-1649 {% for w in audit.findings.tag_format_warnings %} {% endfor %}

bad prefix	count	suggested	sample sessions
`{{ w.bad }}`	{{ w.count }}	{% if w.suggested %}`{{ w.suggested }}`{% else %}unrecognized{% endif %}	{% for s in w.sample_sessions %}`{{ s }}`{% endfor %}

{% else %}

PASS No drift. Baseline {{ audit.baseline_count }} tools tracked.

{% endif %}

Baseline classification breakdown

From .context/audits/orchestrator-mcp-baseline.yaml. T-1166 deprecation annotated in YAML comments — facades that will be removed are not gating targets.

Gated ({{ baseline.gated.count if baseline.gated else 0 }}): {% for t in (baseline.gated.tools if baseline.gated else []) %}{{ t }}{% endfor %}
Mutators ungated ({{ baseline.mutators_ungated.count if baseline.mutators_ungated else 0 }}): {% for t in (baseline.mutators_ungated.tools if baseline.mutators_ungated else []) %}{{ t }}{% endfor %}
Read-only exempt ({{ baseline.readonly_exempt.count if baseline.readonly_exempt else 0 }}): {{ baseline.readonly_exempt.count if baseline.readonly_exempt else 0 }} tools (collapsed)

{% else %}

Audit not yet run. bash agents/audit/orchestrator-mcp-scan.sh or bin/fw audit --section orchestrator.

{% endif %}

Live Sessions ({{ sessions_total }})

{% if sessions_err %}

{{ sessions_err }}

Make sure the TermLink hub is running locally. Routing/orchestration data is live-only; nothing is cached on this page.

{% elif sessions_total == 0 %}

No TermLink sessions registered.

{% else %}

By task-type

{% if task_type_counts %} {% for tt, n in task_type_counts %} {% endfor %}

task-type	count
`{{ tt }}`	{{ n }}

{% else %}

No sessions tagged task-type:<X>.

This is the symptom — orchestrator can route by task-type, but {{ untagged_sessions }} live session(s) carry no such tag. Tracked under T-1643 (Arc B — framework wiring).

{% endif %}

By role

{% if role_counts %} {% for r, n in role_counts %} {% endfor %}

role	count
`{{ r }}`	{{ n }}

{% else %}

No sessions tagged role:<X>.

{% endif %}

Session table ({{ session_rows | length }}{% if session_rows_truncated %} of {{ sessions_total }}, top 50 shown{% endif %})

{% for s in session_rows %} {% endfor %}

name	state	task-type	role	model	task
`{{ s.name }}`	{{ s.state }}	{% for t in s.task_types %}`{{ t }}`{% endfor %}	{% for r in s.roles %}`{{ r }}`{% endfor %}	{% for m in s.models %}`{{ m }}`{% endfor %}	{% for t in s.tasks %}`{{ t }}`{% endfor %}

{% endif %}

Dispatch substrate

Per-model dispatch counts from .context/dispatches.jsonl (CLI parity with fw orchestrator status, T-1788). Each row counts how many times the framework dispatch path actually routed work to that model — the "what got picked" view, complementing the learned-routing "what should win" view below. Synthetic rows (T-stress-*) are excluded from the headline counts.

{% if not substrate.available %}

substrate absent No dispatches.jsonl at {{ substrate.path }} yet. Run a dispatch (fw resolver dispatch or fw termlink dispatch) to seed it.

{% elif substrate.total == 0 and substrate.synthetic_total == 0 %}

no dispatches Substrate file exists but no real dispatches recorded yet.

{% else %}

{{ substrate.total }} real dispatch(es){% if substrate.synthetic_total %}, plus {{ substrate.synthetic_total }} synthetic (T-stress-*, excluded from the breakdown){% endif %}.

{% if substrate.by_model %}

By model

{% for row in substrate.by_model %} {% endfor %}

Model	Dispatches	Share
`{{ row.model }}`	{{ row.count }}	{{ "%.0f%%"\|format(100 * row.count / substrate.total) }}

{% else %}

No model field on any dispatch row yet — substrate rows predate the model-capture path or model wasn't resolved.

{% endif %} {% if substrate.by_task_type %}

By task-type

{% for row in substrate.by_task_type %} {% endfor %}

Task-type	Dispatches	Share
`{{ row.task_type }}`	{{ row.count }}	{{ "%.0f%%"\|format(100 * row.count / substrate.total) }}

{% endif %} {% if substrate.by_worker_kind %}

By worker-kind

{% for row in substrate.by_worker_kind %} {% endfor %}

Worker-kind	Dispatches	Share
`{{ row.worker_kind }}`	{{ row.count }}	{{ "%.0f%%"\|format(100 * row.count / substrate.total) }}

{% endif %} {% endif %}

Outcome quality

Verification pass/fail per task_type from .context/dispatch-outcomes.jsonl (CLI parity with fw orchestrator status --outcomes, T-1749). Latest outcome per dispatch wins (T-1757 dedup rule). Verdict-style events (e.g. escalation-scan-v0.5) are counted in Outcomes but only contribute to passed/failed when they carry verification_passed.

{% if not outcome_quality.available or outcome_quality.total_outcomes == 0 %}

no outcomes No outcome events recorded yet (or dispatch-outcomes.jsonl is absent). Outcomes are written back when dispatch workers exit.

{% else %}

{{ outcome_quality.total_outcomes }} outcome event(s) across {{ outcome_quality.by_task_type|length }} task-type(s).

{% if outcome_quality.by_task_type %} {% for row in outcome_quality.by_task_type %} {% endfor %}

Task-type	Passed	Failed	Total	Pass rate
`{{ row.task_type }}`	{% if row.passed %}{{ row.passed }}{% else %}0{% endif %}	{% if row.failed %}{{ row.failed }}{% else %}0{% endif %}	{{ row.total }}	{% if row.passed + row.failed %}{{ "%.0f%%"\|format(row.pass_rate * 100) }}{% else %}—{% endif %}

{% endif %} {% endif %}

Workflow coverage

Workflow → dispatcher coverage matrix. Each workflow under .context/project/workflows/ declares an optional worker_kind field; the spawn driver (lib/spawn._DISPATCHERS) routes a subset of declarable values. Any workflow that declares a worker_kind without a registered handler is a runtime trap (T-1776 — surfaced as NotImplementedError). T-1798 ratchets this from a runtime trap to an audit-time visibility check; this panel renders the same data on the web.

{% if not workflow_coverage.available %}

unavailable Helper not importable.

{% else %} {% if not workflow_coverage.ok %}

FAIL {% if workflow_coverage.unroutable_workflows %} {{ workflow_coverage.unroutable_workflows | length }} of {{ workflow_coverage.workflows | length }} workflows declare an unroutable worker_kind. {% endif %} {% if workflow_coverage.pi_workflows_missing_provider %} {{ workflow_coverage.pi_workflows_missing_provider | length }} pi workflow(s) missing provider. {% endif %}

{% elif workflow_coverage.warn %}

WARN All {{ workflow_coverage.workflows | length }} workflows route, but {{ workflow_coverage.stale_workflows | length }} workflow(s) stale (no dispatch in 90d) — consider deprecating.

{% else %}

OK All {{ workflow_coverage.workflows | length }} workflows route to a registered dispatcher and pi workflows declare a provider.

{% endif %} {% for w in workflow_coverage.workflows %} {% set is_stale = workflow_coverage.stale_workflows and (w.name in (workflow_coverage.stale_workflows | map(attribute='name') | list)) %} {% endfor %}

Workflow	worker_kind	provider	Routable	Last dispatched
`{{ w.name }}`	{% if w.worker_kind %}`{{ w.worker_kind }}`{% else %}— (interactive){% endif %}	{% if w.provider %} `{{ w.provider }}` {% elif w.worker_kind == "pi" %} missing {% else %} — {% endif %}	{% if not w.worker_kind %} n/a {% elif w.worker_kind in workflow_coverage.routable %} yes {% else %} no {% endif %}	{% if w.last_dispatched %} `{{ w.last_dispatched[:10] }}` {% if w.last_dispatch_task_id %} (`{{ w.last_dispatch_task_id }}`) {% endif %} {% if is_stale %}stale{% endif %} {% else %} never {% if is_stale %}stale{% endif %} {% endif %}

Routable dispatchers: {{ workflow_coverage.routable | join(', ') }}. Declarable but unroutable: {% if workflow_coverage.declarable_but_unroutable %} {{ workflow_coverage.declarable_but_unroutable | join(', ') }} (declared in lib/resolver.VALID_WORKER_KINDS but missing from lib/spawn._DISPATCHERS — declaring any of these in a workflow is a runtime trap). {% else %} none. {% endif %}

{% if workflow_coverage.pi_workflows_missing_provider %}

Missing provider: {{ workflow_coverage.pi_workflows_missing_provider | map(attribute='name') | join(', ') }} (pi workflows without a provider field — lib/spawn._spawn_pi raises SpawnError at runtime; T-1800 surfaces it at audit time).

{% endif %} {% endif %}

Learned routing

Per-task-type model preferences derived from route-cache.json ({{ learned.path }}). Each row shows the model with the highest historical success rate for that task type — the framework dispatch path (T-1669) consults this cache before falling back to FW_DISPATCH_MODEL_FOR_<TYPE> / FW_DISPATCH_MODEL_DEFAULT. Outcomes (success / failure) are recorded back into the same cache after each dispatch's worker exits, so the table grows as the agent dispatches more work.

{% if not learned.available %}

cache absent No route-cache.json at {{ learned.path }} yet. Spawn at least one fw termlink dispatch worker to seed it.

{% elif learned.total_stats == 0 %}

no recordings Cache file exists but no model_stats have been recorded yet. The headline mechanic ("orchestrator picks model based on task_type + historical success rates") needs at least one completed dispatch per task_type to start firing.

{% else %} {% for row in learned.by_task_type %} {% endfor %}

Task-type	Best model	Success rate	Volume	Last used	All candidates
`{{ row.task_type }}`	`{{ row.best.model }}`	{{ "%.0f%%"\|format(row.best.rate * 100) }} ({{ row.best.successes }}/{{ row.best.total }})	{{ row.best.total }}	{{ row.best.last_used or "—" }}	{% for c in row.candidates %} `{{ c.model }}` {{ "%.0f%%"\|format(c.rate * 100) }} {% endfor %}

{{ learned.total_stats }} model/task-type stat(s) total. Resolution order in framework dispatch: explicit --model → route-cache best → env-per-type → env-default → none. Recording uses atomic write + flock for concurrency safety.

{% endif %}

Recent dispatches

Last {{ recent_dispatches|length }} fw termlink dispatch worker(s) from /tmp/tl-dispatch/. model_used and fallback_used are populated at dispatch time by the framework (T-1664) and by /opt/termlink's substrate CLI (scripts/tl-dispatch.sh, /opt/termlink T-1442 commit 143cd870). null means neither path resolved a value (no explicit --model and no FW_DISPATCH_MODEL_DEFAULT / DISPATCH_MODEL_FOR_<TYPE> set).

{% if recent_dispatches %} {% for d in recent_dispatches %} {% endfor %}

Name	Task	Task-type	Model (asked)	Model used	Fallback?	Status	Started
`{{ d.name }}`	{% if d.task %}`{{ d.task }}`{% endif %}	{% if d.task_type %}`{{ d.task_type }}`{% endif %}	{% if d.model %}`{{ d.model }}`{% else %}—{% endif %}	{% if d.model_used %}`{{ d.model_used }}`{% else %}null{% endif %}	{% if d.fallback_used == True %}yes{% elif d.fallback_used == False %}no{% else %}null{% endif %}	{{ d.status }}	{{ d.started }}

{% else %}

No recent dispatches. Run fw termlink dispatch --task T-XXX --name foo --prompt "..." to populate.

{% endif %}

Reconsideration arc

Filed by T-1641 (multi-agent reconsideration of T-1061). Three arcs split the work so reviewers see policy / wiring / drift as distinct decisions, not one mega-arc.

{% for t in arc_tasks %}

{{ t.id }} — {{ t.name }}

type: {{ t.type }} · status: {% if t.completed %}work-completed {% elif t.status == 'started-work' %}{{ t.status }} {% else %}{{ t.status }}{% endif %} · horizon: {{ t.horizon }}

{% endfor %}

Background: T-1641 reconsideration artefact · /review/T-1641 for human decision · 10 worker reports under docs/reports/T-1641-worker-NN-*.md.

Orchestrator Arc

MCP Governance Audit

Live Sessions ({{ sessions_total }})

By task-type

By role

Dispatch substrate

By model

By task-type

By worker-kind

Outcome quality

Workflow coverage

Learned routing

Recent dispatches

Reconsideration arc