roadmap(observability): vendor-neutral execution tracing for agents and DevLoop

Context

mctl already emits or plans the telemetry needed to understand agent execution, but the pieces are tracked as unrelated issues:

mctlhq/mctl-gitops#902 — OpenTelemetry Collector as the mctl telemetry gateway;
mctlhq/mctl-gitops#903 — backend/approach comparison: Traceway, Tempo, Langfuse, Phoenix, SigNoz, Grafana Agent Observability/agento11y;
mctlhq/mctl-agents#195 — end-to-end execution traces across Temporal / Argo / model / MCP / GitHub;
mctlhq/mctl-agent#38 — OpenTelemetry tracing and token-usage metrics in mctl-agent;
mctlhq/mctl-agent#97 — correlate execution traces with Loki, incidents and the mctl MCP surface.

Until now these were only "related work" under mctlhq/.github#18 (agent platform) and mctlhq/.github#48 (agent FinOps). Nothing owned the sequence, so the Collector was never deployed and the backend decision stayed open. This epic owns that sequence.

Goal

Make every mctl execution — DevLoop stage, incident diagnosis, MCP call — observable as one correlated trace, on a vendor-neutral producer boundary:

mctl workloads → OTLP → OpenTelemetry Collector → one or more replaceable backends

Invariants:

OTLP / OpenTelemetry is the only application-side contract. No backend-specific SDK enters mctl-agent or mctl-agents.
Trace identity reuses the execution identity from mctl-agents#196 (execution_id, Temporal workflow/run id, WorkItem, agent/profile version). No second correlation scheme.
Private prompts, completions and Telegram message bodies stay out of telemetry by default.
The trace backend is a view. Durable evidence (mctl-agents#199) and the cost ledger (.github#48) do not depend on trace retention.

Phases

Phase 0 — telemetry gateway

mctl-gitops#902. Deploy the Collector, define the OTLP endpoints workloads use, and pin the resource attributes every producer must set. Fan-out to backends is a Collector concern only.

Phase 1 — backend bake-off

This phase is split at the offline/live boundary.

Phase 1a — comparison harness: mctl-gitops#903 — complete. Frozen rubric, paper screen, representative DevLoop fixtures, generalized exporter/fan-out harness and ADR skeleton are in place.
Phase 1b — live bake-off and decision: mctl-gitops#1280 — open. Deploy the viable candidates, fan out representative traces through the Collector, measure a declared soak window, fill the rubric from observations and close the ADR.

The candidate set is now Traceway, Tempo, Langfuse, Phoenix, SigNoz, plus Grafana Agent Observability / agento11y as an instrumentation/UX approach. Grafana must be evaluated in two distinct shapes:

self-managed agento11y/OTel telemetry → mctl Collector → Tempo/Grafana, preserving the vendor-neutral telemetry boundary;
optional Grafana Cloud Agent Observability for conversation/session, cost/eval and guard UX, explicitly scoring Cloud dependency, privacy/data residency, pricing and portability.

The live bake-off must also prove whether agento11y covers the real claude-agent-sdk DevLoop runtime path, not only interactive Claude Code/Codex-style sessions.

Phase 2 — end-to-end execution traces

mctl-agents#195 and mctl-agent#38. Every DevLoop stage and every mctl-agent diagnosis emits spans carrying the canonical execution identity, including model calls with token usage. This phase is the prerequisite for .github#48 Phase 1 (one DevLoop with cost attribution).

Phase 3 — cross-system correlation

mctl-agent#97. Link traces to Loki log streams, incidents and MCP operations so an operator can move from an alert to the execution that caused it and back.

Current state (2026-10-13)

Verified against GitHub and the live cluster on 2026-10-13. The canonical graph is roadmap/epics/observability.yaml; this section is a snapshot of it.

Work item	Required	Issue	Live evidence
Telemetry gateway, `mctl-gitops#902`	yes	closed 2026-09-11	Collector in `monitoring`. Every producer exports to `otel-collector.monitoring:4318`.
Attribute catalog, `mctl-docs#116`	yes	closed 2026-09-22	`docs/reference/telemetry-attributes.md`
Offline backend evaluation and rubric, `mctl-gitops#903`	yes	closed 2026-09-22	ADR 0001, frozen rubric, fixtures
DevLoop execution tracing, `mctl-agents#195`	yes	closed 2026-10-11	trace `164ac97d8d7264326c406240d263dcc1`
mctl-agent tracing, `mctl-agent#38`	yes	closed 2026-09-23	live since mctl-agent 2.5.0 (gitops#1828), trace `3b54b8b36a8a3334173dc4cad2458e83`
MCP span context, `mctl-api#285`	yes	closed 2026-10-10	`trace_id` in operation results
Cross-system correlation, `mctl-agent#97`	yes	closed 2026-10-13	incident `argo-mctl-agents-implement-dae24112-1791756910` → workflow → trace `431486061c652efb376a860d9d656f25` → ERROR span → Loki. `mctl_get_incident` returns the `trace_id` (mctl-api 4.82.0).
Live Tempo validation and ADR decision, `mctl-gitops#1280`	yes	open	soak in progress, see below

Phase 0, telemetry gateway: done.
Phase 1a, offline selection: done.
Phase 1b, live validation: mctl-gitops#1280 is the only open required item.
- Scope (owner decision 2026-10-10, ADR "Bake-off scope amendment"): the live round is Tempo-only, monolithic on R2. The paper screen already ruled out deploying the other candidates, including the Grafana/agento11y shapes described under Phase 1 above. Under the frozen decision rule, Tempo is the default verdict unless live evidence gives a concrete reason against it.
- Declared soak: from 2026-10-10T11:55Z to 2026-10-23T23:59Z, teardown 2026-10-24.
- Real producers admitted by recorded amendments: the mctl-agents DevLoop (gitops#1788) and mctl-agent (gitops#1860).
- Mid-window record (mctl-gitops#1280, 2026-10-13): 1,840 of 1,840 spans accounted for, no stop rule triggered, 0 leak-check hits.
- Adopting Tempo permanently is a separate owner-gated change after the verdict, not a promotion of the eval deployment.
Phase 2, producer instrumentation and live proofs: done (#195, #38). Prerequisites mctl-gitops#1332 (token redaction) and mctl-agents#196 (execution identity) are closed.
MCP propagation: done (mctl-api#285).
Phase 3, cross-system correlation: done (mctl-agent#97).

Not required by the roadmap, but open: mctl-agent#139 step 3, promoting the catalog attributes that are now shipped. The catalog still marks gen_ai.* as reserved and says no MCTL service emits it.

Remaining critical path

#1280 soak (to 2026-10-23 23:59Z)
   │
   ▼
rubric cells + decision rule → ADR Decision
   │
   ├──► owner-gated adoption change (if Tempo is adopted)
   ▼
teardown of observability-eval (2026-10-24) → verify nothing left running
   │
   ▼
close #1280 → close this epic → roadmap lifecycle

Previous snapshot (2026-09-24), kept for history

Current state (2026-09-24)

The roadmap is in implementation, and the old 2026-09-11 "not started" snapshot is obsolete.

Phase 0 — telemetry gateway: mctl-gitops#902 closed 2026-09-11. The OTel Collector is the vendor-neutral tracing gateway.
Phase 1a — bake-off harness: mctl-gitops#903 closed 2026-09-22. Offline comparison assets, fixtures, fan-out/evaluation harness and ADR skeleton landed. The live/operator portion was deliberately split out.
Phase 1b — live bake-off: mctl-gitops#1280 open. This is the current critical-path work item: deploy viable candidates, run the declared soak, compare real traces, select the architecture in the ADR and remove non-selected candidates. On 2026-09-24 Grafana Agent Observability/agento11y was added to this live comparison as candidate/approach F.
Phase 2a — mctl-agent instrumentation code: mctl-agent#38 closed 2026-09-23 via PR MCP servers: CIMD primary, DCR fallback restricted to the portal callback, full scope for DCR clients #137. Pipeline and gen_ai.* spans plus token metrics are implemented using standard OTEL_* configuration.
Phase 2b — mctl-agent live rollout: mctl-agent#139 open. Release the code, enable otel.enabled in-cluster, verify real spans, then promote telemetry catalog attributes.
Phase 2c — DevLoop end-to-end tracing: mctl-agents#195 open, but the code/offline portion merged in PR #466. Trace propagation Temporal → Argo → pod, model/tool/GitHub/artifact spans, policy events, export-time redaction and failure isolation exist. It remains open for live exported evidence and one real investigation + implementation trace.
Collector correctness blocker: mctl-gitops#1332 open. Current redaction matches the substring token and masks GenAI token-count attributes; fix before using token/cost visibility as bake-off evidence.
Phase 3 — cross-system correlation: mctl-agent#97 open; depends on live end-to-end traces. It will connect incident/GitHub issue → workflow → trace/failed span → Loki and expose trace identity through operational/MCP surfaces.

Current critical path

#1332 fix Collector token redaction
          │
          ├───────────────┐
          ▼               ▼
#1280 live bake-off    #139 agent rollout
          │               │
          └──────┬────────┘
                 ▼
        #195 live DevLoop proof
                 │
                 ▼
          #97 trace ↔ Loki ↔ incident

The immediate roadmap focus is therefore #1280, not creating another backend-comparison epic. Grafana/agento11y is evaluated inside that existing decision path.

Repository ownership

Repository	Ownership
`mctl-gitops`	Collector deployment, backend deployment, retention, dashboards
`mctl-agents`	DevLoop span instrumentation and identity propagation across Temporal / Argo
`mctl-agent`	diagnosis / remediation instrumentation, Loki and incident correlation
`mctl-api`	MCP-side span context and, later, trace links in operation results
`mctl-docs`	telemetry contract, attribute catalog, operator guide

Relationship to other roadmap items

.github#48 agent-finops consumes Phase 2 spans; the cost ledger stays mctl-owned regardless of the backend chosen here.
.github#18 agent-platform supplies the execution identity and profile versions that spans carry.
.github#20 context-platform later attaches ContextSnapshot references to the same traces.

Acceptance criteria

An OpenTelemetry Collector receives OTLP from at least mctl-agents and mctl-agent in the preview cluster. (Since 2026-10-10 / 2026-10-11, see #1280.)
A written ADR selects the backend architecture, with offline comparison evidence from mctl-gitops#903 and live bake-off evidence from mctl-gitops#1280, including the Grafana/agento11y path. (Open: the verdict comes after the soak. The Grafana/agento11y path was screened on paper only, per the 2026-10-10 Tempo-only scope decision.)
One complete DevLoop is visible as a single correlated trace (canonical criterion, roadmap/epics/observability.yaml successCriteria; the roadmap is authoritative). Met by mctl-agents#195, trace 164ac97d8d7264326c406240d263dcc1 in the #1280 Tempo eval backend, a single trace for DevLoop dev-loop-mctlhq-mctl-api-454:
- Traced execution: the mctl-agents-worker Temporal activities (RunActivity:submit_and_wait ×3) with one argo.workflow span per stage (investigate, implement, approve, all Succeeded). Under them, the mctl-agents-investigator and mctl-agents-implementer pods: invoke_agent, chat <model> spans with token usage, execute_tool, and git/GitHub spans (clone, push, issue comment, PR create).
- Reviewer and shepherd activity are not traced, by design. Reviewers run as the claude-review.yml GitHub Action, outside the OTLP producer boundary. The approve and shepherd stages appear in the trace as orchestration spans (the worker's argo.workflow mctl-agents-approve span with its Argo workflow name and phase). Their content is in the existing orchestration and audit evidence: Temporal workflow history, Argo workflow status, mctl-api audit records, and the PR review and merge history on GitHub. No reviewer or shepherd instrumentation will be added to satisfy the wording below.
- Superseded wording (historical, 2026-09)
  
  "One complete DevLoop run is visible as a single trace: investigator → implementer → reviewers → shepherd, with model calls and token usage per span."
  
  This prose predates the deployed tracing design and was broader than the canonical roadmap criterion. It is superseded on 2026-10-13 by the canonical criterion above and is no longer an open acceptance criterion.
One mctl-agent incident diagnosis is visible as a trace linked to the triggering alert and its Loki context. (mctl-agent#97, accepted by the owner 2026-10-12.)
Span attributes follow a documented mctl attribute catalog mapped onto OTel GenAI semantic conventions. (Checked against 1,840 real spans on 2026-10-13: the extra keys are upstream gen_ai.*/error.type, as the catalog requires. The catalog's own status column is stale, see mctl-agent#139.)
No private prompt/completion or message payload is stored in traces by default. (Leak check over all 1,840 stored spans, 0 hits, method on #1280. The poisoned-fixture check is still pending under #1280.)
Replacing the backend requires a Collector configuration change only. (Producers depend only on OpenTelemetry/OTLP and export to the Collector. Tempo is one named Collector exporter.)

Non-goals

Making the trace backend the ledger for cost or the store for execution evidence.
Adding Traceway, Langfuse or any other vendor SDK to application code.
Replacing Loki / VictoriaMetrics / Grafana for logs and platform metrics.
Full-fidelity prompt recording for debugging; that is a separate, opt-in decision.

@mashkovd

Polish 2026-09-11

Read-only pass over the body (no comments yet), the 5 native sub-issues, every issue referenced in text (mctl-agents#196/#197/#198/#199/#60/#149, .github#18/#48/#20), project #1 fields, mctl-gitops/ROADMAP.md, mctl-agents/docs/adr/007, the merged PR history of mctl-gitops, mctl-agents, mctl-agent, mctl-api and mctl-docs, and an org-wide code search for OpenTelemetry / Collector / backend references.

Stale

Nothing stale on the epic itself: all five children are OPEN, none has a closing PR, and all six items (epic + children) sit at Status=Todo in project docs: update org profile status to public/self-service #1. Nothing to untick.
mctl-agents#149, cited by mctl-agents#195 as the isolation prerequisite, is CLOSED (PR mctl-agents#299, 2026-09-04). Informational only; the epic body does not reference it.

Orphan

None. Every sub-issue matches the phase it is assigned to, and every issue the body names either is a sub-issue here or has its proper parent elsewhere (mctl-agents#196/retry(spike): execute #182 governance audit after stale approval grant refusal #199 under .github#18).

Gap

Phase 0 has no contract child. The body says Phase 0 "pins the resource attributes every producer must set" and acceptance criterion 5 requires "a documented mctl attribute catalog mapped onto OTel GenAI semantic conventions", but only the deployment issue mctl-gitops#902 exists. mctl-docs is listed under Repository ownership with no child at all. Suggested child in mctl-docs: docs(observability): mctl telemetry contract and span attribute catalog (execution identity + OTel GenAI semconv mapping).
mctl-api ownership row has no child. "MCP-side span context and, later, trace links in operation results" is Phase 3 work that mctl-agent#97 explicitly leaves to the MCP surface. Suggested child in mctl-api: feat(mcp): propagate OTel span context through MCP calls and return trace_id in operation results.
Phase 1 has no ADR home. mctl-gitops#903 must "close with an ADR", but neither mctl-gitops nor mctl-docs has a docs/adr/ directory; only mctl-agents keeps ADRs (005–008). Where the backend ADR lives needs deciding before #903 can close.

Conflict

mctl-gitops/ROADMAP.md defers tracing. Line 35 lists distributed tracing under "not doing enterprise hardening ahead of demand", and line 149 says OpenTelemetry/tracing comes "only when multi-hop incidents appear that cannot be analysed otherwise". Phase 0 of this epic contradicts the repo-of-record roadmap. Either the trigger is declared met (DevLoop + incident responder are multi-hop today) and ROADMAP.md is amended in the #902 PR, or Phase 0 waits.
mctl-agent#38 targets deleted code. Its proposal instruments internal/diagnosis; that package was removed as dead in mctl-agent#123 (2026-09-05). The live model path is internal/skill/builtin/llm_diagnosis.go, and token metrics would sit next to the existing internal/metrics package. The child body needs a refresh before it is executable.
Circular dependency with mctl-agents#196. This epic says trace identity "reuses the execution identity from mctl-agents#196"; fix(claude-review): tell the reviewer it cannot run checks #196 says it "depends conceptually on spike(mcp-gateway): validate identity, per-tool policy and tenant isolation before migration #195". Neither issue defines execution_id today. ADR-007 explicitly keeps traces and fix(claude-review): tell the reviewer it cannot run checks #196 out of scope, but it does already fix definition_version, profile_version and release_revision on every run, so those three attributes can be pinned in the Phase 0 catalog without waiting on fix(claude-review): tell the reviewer it cannot run checks #196. The remaining identity fields need one declared owner.
mctl-gitops#903 title and candidate list omit SigNoz. The epic and the 2026-09-08 comment on #903 treat SigNoz as candidate E; the child's title and body still list four candidates.
mctl-agent#97 still calls #903 a "Traceway backend POC". #903 has been a five-way comparison since 2026-08-26; the child's reference is out of date.
"mctl already emits … the telemetry" (Context) overstates the baseline. Verified: no OpenTelemetry SDK or OTLP exporter exists in mctl-agent, mctl-agents or mctl-api (mctl-telegram carries go.opentelemetry.io/otel only as an indirect dependency), no Collector manifest exists in mctl-gitops, and mctl-agents#275 exports Temporal SDK metrics to Prometheus only. Starting point is zero traces; recorded in the new Current state section.

Blocker

No item is waiting on a token, credential, repo permission or device. The two items waiting on a human are decisions, listed below, both for @mashkovd.

Next executable step

mctl-gitops#902 — deploy the Collector. It has no upstream dependency, every other child needs an OTLP endpoint to exist (the #903 bake-off fans out behind it, #195/#38 need somewhere to send spans), and it is the only place the ROADMAP.md contradiction can be resolved in code (amend the roadmap in the same PR). The #902 PR should also pin the first resource attributes (service.name, environment, definition_version/profile_version/release_revision from ADR-007) so the docs child has a concrete starting point.

Proposed, needs a decision

ROADMAP.md vs Phase 0: declare the "multi-hop incidents" trigger met and amend mctl-gitops/ROADMAP.md in the #902 PR, or hold Phase 0. Recommendation: declare it met; DevLoop and the incident responder are multi-hop today.
ADR home for the backend decision: create mctl-gitops/docs/adr/ (deployment-owning repo) or put it in mctl-docs. Recommendation: mctl-gitops, next to the manifests it governs.
Create the two missing children (mctl-docs attribute catalog, mctl-api MCP span context) and add them as sub-issues here.
Refresh child bodies: mctl-agent#38 (drop internal/diagnosis, point at internal/skill/builtin/llm_diagnosis.go + internal/metrics), mctl-gitops#903 (title + candidate E SigNoz), mctl-agent#97 (#903 is a comparison, not a Traceway POC).
Identity ownership: state in mctl-agents#196 that it owns the execution_id schema and that spike(mcp-gateway): validate identity, per-tool policy and tenant isolation before migration #195 consumes it, so the circular "depends on" is broken in one direction.

Decision 2026-09-11 on item 5 (identity ownership): mctlhq/mctl-agents#196 owns the execution_id / execution-context schema; #195 and this epic consume it. Dependency edge #195 ← #196 added, so the cycle is broken in that direction. Phase 0 (#902) can proceed with service.name / environment attributes now and adopt the identity attributes when #196 lands.

Decisions applied 2026-09-11

Executed the decisions taken on the Polish 2026-09-11 items 1–4 (item 5 was already decided in the comment above).

ROADMAP.md vs Phase 0 — trigger declared met. Comment on feat(observability): deploy OpenTelemetry Collector as mctl telemetry gateway mctl-gitops#902 (feat(observability): deploy OpenTelemetry Collector as mctl telemetry gateway mctl-gitops#902 (comment)): DevLoop and the incident responder are multi-hop today, so the #902 PR must amend mctl-gitops/ROADMAP.md (lines ~35 and ~149) in the same change and pin the first resource attributes (service.name, environment, definition_version / profile_version / release_revision from ADR-007).
ADR home = mctl-gitops/docs/adr/. Comment on spike(observability): compare Traceway, Tempo, Langfuse, Phoenix, SigNoz and Grafana Agent Observability mctl-gitops#903 (spike(observability): compare Traceway, Tempo, Langfuse, Phoenix, SigNoz and Grafana Agent Observability mctl-gitops#903 (comment)); the directory is created in the PR that lands the ADR. #903 retitled to name all five candidates, and SigNoz moved from the 2026-09-08 comment into the body as candidate E (fan-out diagram, deliverable list and ADR section updated).
Two missing children created and linked as sub-issues here, both in project docs: update org profile status to public/self-service #1 at Status=Todo / Priority=P2:
- Phase 0 contract: docs(observability): mctl telemetry contract and span attribute catalog (execution identity + OTel GenAI semconv mapping) mctl-docs#107 — telemetry contract and span attribute catalog (execution identity + OTel GenAI semconv mapping). Depends on #902 and mctl-agents#196.
- Phase 3: feat(mcp): propagate OTel span context through MCP calls and return trace_id in operation results mctl-api#285 — propagate OTel span context through MCP calls and return trace_id in operation results. Depends on #902, mctl-agents#195/fix(claude-review): tell the reviewer it cannot run checks #196, mctl-agent#97 and mctl-docs#107.
Child bodies refreshed:
- Add OpenTelemetry tracing + token-usage metrics for agent-specific observability mctl-agent#38 — no longer targets the deleted internal/diagnosis; points at internal/skill/builtin/llm_diagnosis.go and internal/metrics; exports over OTLP to the #902 Collector; Depends on: mctlhq/mctl-gitops#902.
- feat(observability): correlate execution traces with Loki, incidents and mctl MCP mctl-agent#97 — #903 is now described as the five-way backend comparison, not a Traceway POC; Depends on: mctlhq/mctl-agents#195.

No code, branches or PRs were created; ROADMAP.md is amended by the #902 PR, not here. Next executable step remains mctlhq/mctl-gitops#902.

Portfolio hygiene — 2026-09-12

Recommended portfolio priority: P1 — active observability foundation.

This snapshot supersedes the stale Current state (2026-09-11) section above.

Live state correction

Phase 0 is done: mctlhq/mctl-gitops#902 (OpenTelemetry Collector telemetry gateway) closed 2026-09-11.
The fixed producer boundary is now real: workloads can target OTLP → Collector without choosing a backend SDK.
Phase 1 mctlhq/mctl-gitops#903 remains open and is the backend-decision track.
Producer instrumentation remains the value-critical work: mctlhq/mctl-agents#195 and mctlhq/mctl-agent#38 must produce a representative real execution trace.

Next executable order

Instrument one representative DevLoop trace through the deployed Collector (#195 + producer side as needed).
Use that same trace for #903 backend comparison rather than benchmarking empty/synthetic infrastructure only.
Then land mctlhq/mctl-agent#97 cross-system incident/Loki/MCP correlation.

Portfolio rule

Do not make the backend bake-off the critical path to emitting traces. OTLP/Collector is already the contract; backend choice must remain replaceable and must not block producer instrumentation.

Producer-side input from mctlhq/mctl-telegram#617 (Phase 3 of the enterprise MCP roadmap .github#35), shipped in mctl-telegram 0.66.0.

tg.mctl.ai now records how every MCP call arrived, in audit_logs and mirrored to its structured log line: edge_request_id (the Cloudflare ray), edge_route (portal/direct), mcp_method, mcp_name, protocol_version. Captured once per request in internal/edgectx and read from ctx at the audit write, so there is a single capture point for whatever consumes it next.

The OTel half of #617 is deliberately not done there and belongs to this epic. mctl-telegram has no tracer — opentelemetry is an indirect dependency only — and adding one locally would have created the second correlation scheme this epic's invariants forbid. When Phase 0's Collector lands and a tracer reaches this service, the span attributes are already settled and read the same values the audit row carries:

span attribute	source column
`mctl.edge.request_id`	`edge_request_id`
`mctl.edge.route`	`edge_route`
`mcp.method` / `mcp.name`	`mcp_method` / `mcp_name`
`mcp.protocol_version`	`protocol_version`

Three measured facts worth having in this epic's context before anyone designs trace correlation across the Cloudflare boundary:

Cf-Ray arrives on both routes — the zone is proxied, so a direct consumer call gets one too. It identifies a request, not a route; the discriminator is Cf-Worker.
One portal tool call is two upstream requests (server/discover, then tools/call), each with its own ray. A per-call span must expect the pair.
The client→mcp.mctl.ai leg cannot be tied to the subrequest leg today. Nothing in the arriving request carries the client-leg ray, and Cloudflare's per-request dataset is unavailable on this zone's Free plan (parentRayName absent from the schema, rayName entitlement-denied, adaptive groups empty, no Logpush). So cross-layer correlation with Cloudflare evidence is by principal and time; a paid plan would change that, which makes it a cost decision rather than an engineering one.

Also worth carrying over: these fields are headers as received — evidence for reading a trail, never an authorization input. A direct caller cannot forge edge_route on this zone (Cloudflare strips client-supplied Cf-*, measured), but that is a property of the current Cloudflare configuration rather than of mctl.

Full measurements: mctl-telegram#617 — header table, worked example, forgery test, ray result.

Final state — completed

Closing this epic after the canonical roadmap reached completion:

8/8 required work items complete;
0 blocked, 0 unknown;
RoadmapHealth healthy, drift=0, error=0;
live backend validation #1280 completed with ADOPT TEMPO / mctl-native only;
eval teardown #2073 merged;
permanent Tempo adoption #2072 merged;
roadmap/epics/observability.yaml lifecycle changed from active to completed in chore(roadmap): complete observability epic #296 (merge 69d0f1e).

The informational HierarchyUnexpectedChild diagnostic for mctlhq/mctl-docs#107 is non-blocking housekeeping and is not part of this epic's completion gate.

Navigation Menu

Repository navigation

roadmap(observability): vendor-neutral execution tracing for agents and DevLoop #55

Context

Goal

Phases

Phase 0 — telemetry gateway

Phase 1 — backend bake-off

Phase 2 — end-to-end execution traces

Phase 3 — cross-system correlation

Current state (2026-10-13)

Remaining critical path

Current state (2026-09-24)

Current critical path

Repository ownership

Relationship to other roadmap items

Acceptance criteria

Non-goals

Polish 2026-09-11

Stale

Orphan

Gap

Conflict

Blocker

Next executable step

Proposed, needs a decision

Decisions applied 2026-09-11

Portfolio hygiene — 2026-09-12

Live state correction

Next executable order

Portfolio rule

Final state — completed

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

roadmap(observability): vendor-neutral execution tracing for agents and DevLoop #55

Description

Context

Goal

Phases

Phase 0 — telemetry gateway

Phase 1 — backend bake-off

Phase 2 — end-to-end execution traces

Phase 3 — cross-system correlation

Current state (2026-10-13)

Remaining critical path

Current state (2026-09-24)

Current critical path

Repository ownership

Relationship to other roadmap items

Acceptance criteria

Non-goals

Activity

mashkovd commented on Sep 10, 2026

Polish 2026-09-11

Stale

Orphan

Gap

Conflict

Blocker

Next executable step

Proposed, needs a decision

mashkovd commented on Sep 11, 2026

mashkovd commented on Sep 11, 2026

Decisions applied 2026-09-11

mashkovd commented on Sep 11, 2026

Portfolio hygiene — 2026-09-12

Live state correction

Next executable order

Portfolio rule

mashkovd commented on Sep 12, 2026

mashkovd commented on Nov 1, 2026

Final state — completed

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions