Tracking: monitor-local-agents — phase-by-phase ownership map and review-support offer

[yashksaini-coder]

The monitor-local-agents initiative now has 23 open issues across 6 phases, with several contributors working in adjacent areas. I'm posting this as a single coordination thread so we can keep dependencies and ownership visible without scattering the conversation across 23 issue threads.

Where things stand

Phase 0 — Foundation

The Click command group skeleton (#1486) is up as PR #1532, CI green, awaiting review. Once it lands, the rest of the chain has a CLI surface to attach to.

Phase 1 — Golden signals MVP

@VibhorGautam is on #1487 (the AgentRegistry), which is the actual prerequisite for the visible dashboard. The probe sampler (#1489), the wiring layer (#1490), and the active/idle/stuck heuristic (#1491) are mine — I'm planning to ship them in that order over the next two weeks. /agents kill (#1492) and /agents trace (#1493) wait until the registry shape is final.

Phase 0/1 cross-coordination: @muddlebee and I are co-assigned on #1488 (rendering rows from the registry). I left a coordination note on the issue proposing a clean split — picking up either the table-rendering shape or the slash-command wiring, whichever they'd rather not take.

Phase 2 — Quality & cost

I picked up #1495 (the TokenMeter base interface plus the Claude Code parser) because it's independent of the registry and can ship in parallel with #1487. The other three Phase-2 tickets — #1494, #1496, #1497 — are leaf-independent and a good pick for new contributors.

Phases 3–5

@yug105 is on #1506 (Phase 3, LocalProcessIntrospectTool). @Mohdtalibakhtar is on #1502 (Phase 4, conflict map). #1501 (orphan sweep) is mine but blocked on the registry. The remaining Phase 3/4/5 tickets are open — see the list at the end of this post.

Review-support I'm offering

If you're working on a ticket adjacent to one of mine, I'm happy to:

• Same-day review on #1487 drafts. Four of my tickets land on the AgentRecord shape that's chosen there, so I have a real interest in catching any issues before merge rather than after.
• Compare notes on the registry-consuming surfaces (#1488, #1492, #1493, #1501) so we don't all assume different lookup APIs.
• Pre-PR shape check for anything in app/agents/ or the /agents slash command surface — ping me on the issue and I'll read it the same day.

Tickets currently open for new contributors

Leaf-independent (start any time, no in-flight dependency):

• #1494 — agents.yaml config loader and /agents budget
• #1496 — Loop-detection rolling-shingle module
• #1497 — Error/retry rate detector
• #1498 — REPL Ctrl-C intervention counter
• #1499 — GuardrailStream chunked adapter
• #1500 — Filesystem blast-radius watcher
• #1503 — /agents claim and /agents release
• #1504 — /agents graph dependency rendering

Larger-scope (probably want maintainer scoping first):

• #1505 — Shared context bus
• #1507 and #1508 — Phase 5 incident response

Cadence

I'll keep this thread updated as PRs open and merge on my side. If you're working on a ticket in the table above and I haven't pinged you yet, please feel free to reply here with where you are — even a one-line "started" or "blocked on X" helps everyone else plan around it.

cc @davincios — flagging in case any of this conflicts with how you'd prefer the coordination to go. Happy to restructure the thread if it's not the shape you want.

[yashksaini-coder]

Progress update — afternoon of 2026-05-07

Quick status since the OP. Noting the merge of the foundation ticket and where things stand on the four PRs in flight.

Phase 0 — Foundation

• #1486 is merged (PR #1532 by @muddlebee at 13:59Z). The agents Click command group now lives on main, which means the rest of the chain has a CLI surface to attach to. Phase 0 done.

Phase 1

• #1487 (registry) — still in progress with @VibhorGautam, no public update. Four of my own tickets (#1488, #1492, #1493, #1501) are sitting in the waiting room until the registry shape is final.

• #1489 (psutil probe) — PR #1565 at commit f552c946. All review feedback addressed: Greptile flagged psutil.AccessDenied could escape from probe() (it can on macOS / restricted Linux for cross-user processes); fixed by widening the inner except. Also tightened _BYTES_PER_MIB naming and a lazy hasattr fallback for net_connections per Copilot. CI green, ready for maintainer review.

Phase 2

• #1495 (TokenMeter) — PR #1567 at 7a1b535a. Greptile 5/5 confirmed. The interesting catch was the original parser silently overstating tokens/min by ~50% in any multi-turn session — the result event in Claude Code's stream-json carries cumulative totals, not a per-turn final repeat. Switched from regex-on-text to JSON-per-line filtered to type == "assistant" events; fixture-test expectation went from 1142 → 756 (the actual sum across the three assistant events). Same change also kills the false-positive risk where a quoted-key form ("input_tokens": 5000) embedded in assistant prose would have been counted.

Phase 3 cross-cutting (bonus)

• #1377 (/privacy controls + history redaction) — PR #1387. Greptile is at 5/5. @muddlebee left two pieces of feedback that I addressed in efd66e87: (1) moved the redaction patterns into a dedicated app/cli/interactive_shell/redaction_rules.py constants module so they're easier to audit and extend, and (2) scanned app/services/ + app/integrations/ for credential env-var names (44 distinct ones) and expanded the ruleset 12 → 22 patterns covering Google/GitHub OAuth/GitLab/OpenRouter/HuggingFace/Stripe-restricted, DSN-style connection strings with embedded user:password@, and generic <NAME>_(API_KEY|TOKEN|PASSWORD|SECRET)=value env-style fallbacks for the long tail. 40 new tests, all 442 existing tests/cli/interactive_shell/ tests still pass. The commit is on my fork; pushing it to upstream is the only thing left.

Coordination

• Looking for review on: PR #1565 and PR #1567. Both have CI green and review feedback fully addressed.
• Still happy to review #1487 drafts the same day — the AgentRecord shape lands in four downstream tickets I'm assigned to.
• Still open for new contributors (no in-flight dependency): #1494, #1496, #1497, #1498, #1499, #1500, #1503, #1504.

Minor lessons from the round

• Greptile re-reviews are not opt-out free — they edit the same summary comment in place rather than posting a new one, so chasing "is this score current?" requires checking the Last reviewed commit: line in the body, not just the timestamp.
• Greptile's re-trigger endpoint accepts requests (HTTP 307) but de-dupes within a short window. Multiple re-triggers don't help; only a new commit SHA forces a guaranteed fresh review.
• GitHub's push-protection secret scanner flags test fixtures that look too realistic (e.g. a glpat-abc123XYZ_-defGHI456 synthetic token). Repeated-literal placeholders ("FAKEFAKE" * 4) avoid the false positive while still satisfying length-based rule patterns.

cc @muddlebee @VibhorGautam @yug105 @Mohdtalibakhtar — flagging the merge of #1486 since it's relevant to anyone whose ticket will eventually land on top of it. cc @davincios — happy to restructure this thread if you'd prefer a different shape.

[yashksaini-coder]

Progress update — 2026-05-08 morning

Big movement overnight. Posting an update because the dependency picture for the rest of this initiative has changed substantially.

What landed

#	Title	Merged via	When
#1486	Register the agents Click command group	PR #1532	2026-05-07T13:59Z
#1487	AgentRecord + AgentRegistry with JSONL persistence	PR #1600 (@VibhorGautam)	2026-05-08
#1495	TokenMeter base + Claude Code parser	PR #1567	2026-05-08T05:59Z
#1499	GuardrailStream chunked adapter	(PR by @nawneet77)	2026-05-07
#1502	Conflict map + /agents conflicts	(PR by @Mohdtalibakhtar)	2026-05-07

That's 5 of 23 issues now closed. Phases 0, 2, 3, and 4 each have at least one piece in place; Phase 1's foundation (registry + Click group) is fully complete.

What this unblocks

@VibhorGautam's #1487 merge is the highest-leverage event since #1486's. Four of my queued tickets — #1488 (rendering), #1492 (kill), #1493 (trace), #1501 (orphan sweep) — all gated on the AgentRecord shape, and they can now move. Combined with #1490 and #1491 (already unblocked), that's six PRs of mine in the queue.

What's still in flight

PR #1565 (#1489 — psutil per-PID probe) is the only remaining open PR on my side. It needed a rebase this morning because #1567 (TokenMeter) and #1565 both added app/agents/__init__.py. Resolved by taking main's version (which had grown to include AgentRecord/AgentRegistry from #1487 too) and replaying the two probe commits on top. Branch is at c4cdcca0, all checks green, Greptile re-triggered.

@muddlebee — flagging that the conflict you asked me to fix is resolved and pushed.

Sequencing for the next two weeks

Today / next 24h:
  #1565 lands once Greptile catches up

This week:
  #1490 PR — wire probe into REPL background task (needs #1565)
  #1491 PR — active/idle/stuck heuristic (needs #1490's wired data)
  #1488 PR — slash-command rendering (with @muddlebee, needs registry — already in main)

Next week:
  #1492 PR — /agents kill <pid>
  #1493 PR — /agents trace <pid> live tail
  #1501 PR — startup orphan / stale-lockfile sweep

That's six PRs over ~2 weeks at sustainable pace — each an additive slice of the dashboard surface. After that, only Phase 5 incident-response tickets (#1507, #1508) and a few independent quality-and-cost modules remain.

Coordination

• @muddlebee — open question on #1488 (we're co-assigned). The cleanest split feels like: one of us takes the Rich/textual rendering shape for the registry rows, the other wires the slash command into the REPL command_registry/. Happy to take whichever you prefer or step aside if you want to lead solo. Either way, let's settle the split before either of us opens a PR — would rather coordinate than rebase over each other.
• New contributors @nawneet77, @psyberck — welcome! If you're working on adjacent surfaces and want a pre-PR shape check, ping me on the issue.
• Still happy to review anything in app/agents/ or the /agents slash command surface same-day.

Tickets still open for new contributors

Leaf-independent (no in-flight dependency, ready to start):

• #1494 — agents.yaml config loader and /agents budget
• #1496 — Loop-detection rolling-shingle module
• #1497 — Error/retry rate detector
• #1500 — Filesystem blast-radius watcher
• #1503 — /agents claim and /agents release

Larger-scope (probably want maintainer scoping first):

• #1505 — Shared context bus
• #1507 and #1508 — Phase 5 incident response

Lessons from this round

• Stack-of-PRs trade-off: parallel PRs let me ship 3 simultaneously without rebasing each one. But the moment the first one merges, the next has to absorb the conflict surface. PR feat(agents): per-PID probe sampler with psutil (#1489) #1565 demonstrated this — both feat(agents): TokenMeter base + Claude Code parser (#1495) #1567 and feat(agents): per-PID probe sampler with psutil (#1489) #1565 added app/agents/__init__.py, so the merge of the former forced a rebase of the latter. Useful to plan for one rebase round per parallel-PR-batch.
• Greptile editing in place: worth knowing across the team — Greptile mutates summary comments rather than appending. The score in the body changes over time on the same comment ID. Only the Last reviewed commit: line is a reliable pointer to what was actually scored.
• GitHub push-protection on test fixtures: GitHub's secret scanner flags realistic-looking tokens in code, including in tests. For fixture credentials, use repeated-literal placeholders ("FAKEFAKE" * 4) that satisfy length-based regex rules without triggering the scanner.

cc @davincios — happy to keep this thread updated weekly or when something material changes; let me know if you'd prefer a different cadence or shape.

[yashksaini-coder]

Update — all four of my opened PRs merged

PR #1565 (#1489 — psutil per-PID probe) just landed. That closes out the batch I had in flight.

#	Issue	PR	Merged
1	#1486 — Click group skeleton	#1532	2026-05-07T13:59Z
2	#1377 — privacy / history redaction	#1387	2026-05-08T05:37Z
3	#1495 — TokenMeter	#1567	2026-05-08T05:59Z
4	#1489 — psutil probe	#1565	2026-05-08T06:17Z

Thanks @muddlebee for the reviews and merges, @davincios for the decomposition that made this shippable in slices, and @VibhorGautam for #1487 (AgentRegistry) which made half of these PRs possible.

Initiative-wide progress

6 of 23 issues closed — about 26% of the monitor-local-agents scope:

• #1486 — Click group (me)
• #1487 — AgentRegistry (@VibhorGautam)
• #1489 — psutil probe (me)
• #1495 — TokenMeter (me)
• #1499 — GuardrailStream adapter (@nawneet77)
• #1502 — Conflict map / /agents conflicts (@Mohdtalibakhtar)

Phase 0 is fully complete. Phase 1 has the registry + Click group + probe — only the wiring (#1490, #1491) and the slash-command-table consumers (#1488, #1492, #1493) remain. Phase 2 has the TokenMeter shipped; the consumer-side wiring rides on top of #1490. Phase 3 is partly underway (#1499 done, #1498 in progress with @Mohdtalibakhtar). Phase 4 has @Mohdtalibakhtar's #1502 done and @psyberck on #1504. Phase 5 still untouched.

What I'm picking up next

In order, all already unblocked:

1. #1490 — wire the probe into the REPL background task. Direct continuation of Add psutil dependency and app/agents/probe.py per-PID sampler #1489. Single most important next ticket because it's what makes the /agents table show real numbers.
2. #1491 — active/idle/stuck heuristic on top of wired probe data. Pure function over a window of probe samples; depends on Wire the probe into the REPL background task and populate /agents columns #1490.
3. #1488 — /agents slash command rendering rows from the registry. Co-assigned with @muddlebee — let's settle the split before either of us opens the PR. Suggestion in my earlier reply: one half is the Rich/textual rendering shape, the other is wiring /agents into app/cli/interactive_shell/command_registry/. Open to either.
4. #1492 + #1493 — /agents kill <pid> and /agents trace <pid> live tail. Both depend on Add /agents slash command rendering registered rows #1488's command surface. Likely a stacked PR pair.
5. #1501 — startup orphan / stale-lockfile sweep against the registry. Ships independently once the registry's persistence semantics are stable (they are, as of Add AgentRecord dataclass and AgentRegistry with JSONL persistence #1487).

That's 5 more PRs, plausibly over ~10 days at sustainable pace.

Still open for new contributors

If anyone joining wants a leaf-independent ticket (no in-flight dependency), these are the cleanest picks:

• #1494 — agents.yaml config loader and /agents budget
• #1496 — Loop-detection rolling-shingle module
• #1497 — Error/retry rate detector
• #1500 — Filesystem blast-radius watcher
• #1503 — /agents claim and /agents release for branch ownership

Larger-scope (probably want maintainer scoping discussion first):

• #1505 — Shared context bus
• #1507 and #1508 — Phase 5 incident response (depends on Phases 1-3 actually working end to end, which we're close to)

Quick reflection on what shipped

• Click group + registry + probe is enough to show something in the /agents table — the cells will be blank until Wire the probe into the REPL background task and populate /agents columns #1490 wires the probe samples in, but the table renders, the slash command resolves, the data shape exists.
• TokenMeter independently catches an interesting class of dashboard bug — the result event in Claude Code's stream-json was overstating per-session totals by ~50% under the original regex. Worth flagging to anyone working on similar collectors for other CLIs (Codex, Cursor, Aider) to check that you're filtering to per-turn events, not session-summary events.
• Privacy redaction got 22 patterns (up from 12) covering most of the credentials the project actually handles. The structure (constants module + grouped categories + generic env-style fallbacks) is extensible — adding a new pattern is now a one-line tuple append plus a parametrized test case.

I'll keep this thread updated as the next batch of PRs opens. Realistic next checkpoint: #1490 PR open within 24-48h, #1491 right after.

cc @davincios @muddlebee — appreciate the fast review cycle this round.

[yashksaini-coder]

Update — narrowing scope to the two open PRs

Quick note for transparency. I'd self-assigned a downstream queue (#1490, #1491, #1492, #1493) earlier in the week as part of laying out a sequenced shipping plan. With four PRs already merged this round and two more in review (#1624, #1627), the queue had grown longer than what I'm actually going to ship next.

I've removed those self-assignments and deleted the speculative claim comment on #1490. The four tickets are now back in the available pool:

• #1490 — Wire the probe into the REPL background task and populate /agents columns
• #1491 — Status heuristic: active / idle / stuck
• #1492 — /agents kill <pid> slash command
• #1493 — /agents trace <pid> live tail view

All four are now genuinely unblocked — AgentRegistry (#1487) and the per-PID probe (#1489) are in main, so a new contributor can pick any of them up directly. #1490 is the most leveraged of the four since landing it makes the /agents table show real cpu% / uptime cells; #1492 and #1493 attach to the slash-command surface from PR #1627 once that lands.

My active scope is now just:

• PR #1624 — startup orphan / stale-lockfile sweep (closes Startup orphan / stale lockfile sweep #1501). Awaiting maintainer review on the rebased branch with all bot feedback addressed.
• PR #1627 — /agents slash command rendering registered rows (closes Add /agents slash command rendering registered rows #1488, co-assigned with @muddlebee). Awaiting review.

Both review-clean. Not picking up new tickets until those land.

cc @davincios — releasing the speculative queue per your "redistribute to community" pattern. Happy to pick up something concrete again once the open PRs are merged and there's clear demand.