context-garden
Recorded snapshot · 2026-09-06T02:34:47+00:00. Main regions render the supplied garden export. State atlas below is simulated. Task, run and Inbox links show live-app destinations; they require a running app. Clocks replay from capture time; this is not a live connection.

Plate V · context-garden

Work in view.

What is moving, what comes next, and what it adds up to.

September 6, 2026
5 run records · 3 without a process

Now

2 processes present · 3 missing
Ⅱ Merge held

Hold untrusted config changes before live reload

the automated review verdict is request_changes, not approve

View decision · Open in Inbox

Process missing · revise

Process not found · run record still running

`garden dispatch <id>` (CLI) allows dispatching a draft directly

harness not recorded · model not recorded · tier not recorded

No assistant text recorded.

No assistant text recorded.

Elapsed since start · activity unconfirmed

Typical not established

— · not reported yet

Open run · Reconciliation timing unknown; inspect the run for its latest status.

Process missing · revise

Process not found · run record still running

Hold untrusted config changes before live reload

harness not recorded · model not recorded · tier not recorded

No assistant text recorded.

No assistant text recorded.

Elapsed since start · activity unconfirmed

Typical not established

— · not reported yet

Open run · Reconciliation timing unknown; inspect the run for its latest status.

Process missing · revise

Process not found · run record still running

Isolate planner execution from operator state

harness not recorded · model not recorded · tier not recorded

No assistant text recorded.

No assistant text recorded.

Elapsed since start · activity unconfirmed

Typical not established

— · not reported yet

Open run · Reconciliation timing unknown; inspect the run for its latest status.

Process status comes from the capture. “Implemented” in an assistant message does not finish a run.

Next

at the next opening

Each queue moves on its own.

Workers · 5 / 5 recorded occupied

  1. Worker temp files live on disk and are pruned: TMPDIR under the work root, per-run cleanup at reap, and a sweep of finished tasks' worktree venvs and caches

    work · priority 1

  2. The task page shows the decision a worker's no-change or question report needs: the same card and actions as the Inbox, right where the notification sends you

    work · priority 1

  3. Inbox decision cards lay out at full width: the text column no longer collapses to one word per line and the action buttons no longer overlap the evidence list

    work · priority 1

25 more in dispatch order

  1. Design Now 2: astra's take on the Now page, information architecture, visual system, motion, and a static mock of every state

    work · priority 1

  2. UI changes are reviewed against rendered pages: a template or style change captures the affected pages as screenshots at two widths, the reviewer and the personas read them, and the walkthrough uses the same capture

    work · priority 1

  3. A dispatch that fails before its process starts closes the run record at once, and the orphan sweep closes any running record with no live process

    work · priority 1

  4. The task page renders a trial in progress: the trial panel treats winner, scores and PRs as optional, and a test renders a task mid-trial with a failed contender

    work · priority 1

  5. The web app serves design documents, mocks and run captures: /design/<file> for the product's docs/design and a run page link to each capture

    work · priority 1

  6. Onboarding skill and command: analyse an existing project and its environment to create a garden product, principles, setup config and a first phase

    work · priority 2

  7. The retro captures its own walkthrough, and hand merges and tick duration are in garden metrics and the rail

    work · priority 2

  8. CG-217 merged with a criterion marked 'no evidence given' and CG-158 with a plac

    work · priority 2

  9. The built-in economy, balanced and fast stops hardcode Claude model ids and mode

    work · priority 2

  10. A brief never ships with an empty or unresolved reading list, and a revise brief restates the criteria and the concrete blocker

    work · priority 2

  11. Docs match the mechanism: the scaffolded operate skill, design.md and roadmap.md non-goals, and the architecture module map

    work · priority 2

  12. The retro's Numbers section reads the operator ledger where the owner keeps it, and reports spend and share

    work · priority 2

  13. Retro questions are deduplicated across reconcile runs, and a second judge can run with task filing off

    work · priority 2

  14. Spike: OpenRouter harness shape and adapter CLI

    work · priority 2

  15. parse_result accepts the GARDEN_RESULT marker wrapped in markdown emphasis or code and a JSON payload that spans lines

    work · priority 2

  16. redispatch kills the superseded worker, and pin runs the canary, installs and restarts after a tick

    work · priority 3

  17. A draft's acceptance criteria and reading list can be edited inline on the task page

    work · priority 3

  18. Planning sequences dependent tasks and inlines retro evidence into the brief

    work · priority 3

  19. One vocabulary and one place for each fact across rail, Config, CLI and Inbox

    work · priority 3

  20. The walkthrough renderer skips hidden elements and attributes, and check runs retry once on a signal exit

    work · priority 3

  21. Update docs: docs/architecture.md

    work · priority 3

  22. Update docs: docs/worker-protocol.md

    work · priority 3

  23. Update docs: docs/design.md

    work · priority 3

  24. Persona runs are recorded per phase and the retro validates the run id it reads

    work · priority 4

  25. Small follow-ups: retro_question kind, backlog noscript, CI flake, ambient config dir in tests, a second opinion for self products

    work · priority 4

Merges

No eligible merge candidates.

Hold untrusted config changes before live reload

review round 2 / 4 · CI pending

Reviews · 0 / 3 slots occupied

Hold untrusted config changes before live reload

Exported reason: no review slot (0 of 3 busy)

Queue reason conflicts with available capacity. Reconciliation timing unknown. View decision for the current review state.

Last tick · 2026-09-06T02:33:04+00:00
Next tick not reported
Dispatch pause state not supplied

Progress in context

Where we are

prickly poppy, botanical plate by Thomé, 1885

phase-05 · PLATE V

Papaver argemone

10 / 45

tasks done

in leaf in leaf · 22%

A garden in someone else’s environment.

Adoption, measured by the cost of accepted work.

  • ○Onboardingnot started · 0 / 1 linked tasks
  • ○Any model, at the right pricenot started · 0 / 1 linked tasks
  • ○Shared quotasnot started · 0 / 1 linked tasks
  • ○Any machinenot started · 0 / 1 linked tasks
  • →What the phase-04 retro addsin flight · 3 / 4 linked tasks

Retro verdict · not yet recorded
Cancelled / won't-do breakdown not supplied

Dryopteris filix-mas

phase-04

61 / 62 done · plate IV

Retro: reopen · accepted

Frozen

Paeonia mascula

phase-06

0 / 1 done · plate VI

No retro verdict

Open

Earlier specimens · closed phases

Digitalis purpurea
Plate III
30 done
Pisum sativum
Plate I
19 done
Rubus fruticosus agg.
Plate II
90 done

Thomé, Flora von Deutschland, 1885.
Plate provenance: static/plates/SOURCES.md.

Outcomes, not activity alone

The last period

Last hour · 2026-09-06T01:34:47+00:00 → 2026-09-06T02:34:47+00:00

8 merged in this window

$28.45 recorded window spend

All recorded activities in the selected window; price completeness unknown.

Accepted-task cost · separate lifetime measure

$4.53 / accepted task, as exported.

Intended measure: mean lifetime cost of tasks accepted in this window. Exported merged cohort: 8 tasks. Priced denominator and lifetime-cost coverage not supplied; this figure is unverified. It is not window spend divided by merges.

0% first-pass approval

n=2 reviewed tasks (export cohort) · 24 completed runs

Run completions by exported hour bucket: 9, 15.
Accepted-task throughput timestamps not supplied.

1 recorded configuration annotations

2026-09-06T02:01:11+00:00 · config_reloaded · harness, review

Scroll to compare models →

Cost per accepted tasklower is better · n = accepted tasks · exported measurements
Difficultyclaude-sonnet-5gpt-5.6-terraclaude-opus-4-8claude-fable-5-1
easy$3.00=n=2 †—n=0—n=0—n=0
medium$5.22↓n=1 †$3.08↑n=3—n=0—n=0
hard—n=0—n=0$7.27=n=2 †—n=0
First-pass approvalhigher is better · n = reviewed tasks (export cohort) · exported measurements
Difficultyclaude-sonnet-5gpt-5.6-terraclaude-opus-4-8claude-fable-5-1
easy0%=n=1 †—n=0—n=0—n=0
medium—n=0—n=0—n=0—n=0
hard—n=0—n=0—n=00%=n=1 †
Work-run costlower is better · n = work runs · exported measurements
Difficultyclaude-sonnet-5gpt-5.6-terraclaude-opus-4-8claude-fable-5-1
easy—n=0—n=0—n=0—n=0
medium$0.62↓n=2 †$0.53↑n=2 †—n=0—n=0
hard—n=0—n=0—n=0$11.32=n=2 †
Revise roundslower is better · n = accepted tasks · exported measurements
Difficultyclaude-sonnet-5gpt-5.6-terraclaude-opus-4-8claude-fable-5-1
easy0.0=n=2 †—n=0—n=0—n=0
medium1.0↑n=1 †1.3↓n=3—n=0—n=0
hard—n=0—n=00.0=n=2 †—n=0
Median lead timelower is better · n = accepted tasks with lead times · exported measurements
Difficultyclaude-sonnet-5gpt-5.6-terraclaude-opus-4-8claude-fable-5-1
easy94 min=n=2 †—n=0—n=0—n=0
medium115 min↓n=1 †97 min↑n=3—n=0—n=0
hard—n=0—n=0119 min=n=2 †—n=0

↑ best · ↓ worst · = tied or only value · † n < 3, provisional
Grounds compare within a row. Colour is never the only signal.

A task using several models appears in each. Columns do not add up. Exported cohorts are preserved; the build’s shared computation contract is in the design document.

Recorded spend by activity

check
$0.00
other
$22.65
rebase
$0.49
review
$3.01
revise
$2.30

Runs and hand steps

claude:claude-fable-5-1
2 runs
garden:review
5 runs
claude:claude-sonnet-5
3 runs
codex:gpt-5.6-terra
2 runs
codex:gpt-5.6-luna
1 runs
garden:rebase
2 runs
garden:check
9 runs
Hand steps
3 recorded
Manual merges
not supplied
Recorded hand-step kinds
dispatch_paused
1
dispatch_resumed
2

Today (UTC) · 2026-09-06T00:00:00+00:00 → 2026-09-06T02:34:47+00:00

10 merged in this window

$88.90 recorded window spend

All recorded activities in the selected window; price completeness unknown.

Accepted-task cost · separate lifetime measure

$4.41 / accepted task, as exported.

Intended measure: mean lifetime cost of tasks accepted in this window. Exported merged cohort: 10 tasks. Priced denominator and lifetime-cost coverage not supplied; this figure is unverified. It is not window spend divided by merges.

42% first-pass approval

n=12 reviewed tasks (export cohort) · 101 completed runs

Run completions by exported hour bucket: 70, 16, 15.
Accepted-task throughput timestamps not supplied.

1 recorded configuration annotations

2026-09-06T02:01:11+00:00 · config_reloaded · harness, review

Scroll to compare models →

Cost per accepted tasklower is better · n = accepted tasks · exported measurements
Difficultyclaude-sonnet-5gpt-5.6-terraclaude-opus-4-8claude-fable-5-1gpt-6-astra
easy$2.50=n=3—n=0—n=0—n=0—n=0
medium$5.80↓n=2 †$3.08↑n=3—n=0—n=0—n=0
hard—n=0—n=0$7.27=n=2 †—n=0—n=0
First-pass approvalhigher is better · n = reviewed tasks (export cohort) · exported measurements
Difficultyclaude-sonnet-5gpt-5.6-terraclaude-opus-4-8claude-fable-5-1gpt-6-astra
easy67%=n=3—n=0—n=0—n=0—n=0
medium50%↑n=2 †0%↓n=4—n=0—n=0—n=0
hard—n=0—n=0100%↑n=2 †0%↓n=1 †—n=0
Work-run costlower is better · n = work runs · exported measurements
Difficultyclaude-sonnet-5gpt-5.6-terraclaude-opus-4-8claude-fable-5-1gpt-6-astra
easy$1.94=n=3—n=0—n=0—n=0—n=0
medium$3.32↓n=5$0.72↑n=10—n=0—n=0—n=0
hard—n=0—n=0$6.24n=2 †$11.32↓n=2 †$5.98↑n=1 †
Revise roundslower is better · n = accepted tasks · exported measurements
Difficultyclaude-sonnet-5gpt-5.6-terraclaude-opus-4-8claude-fable-5-1gpt-6-astra
easy0.0=n=3—n=0—n=0—n=0—n=0
medium1.0↑n=2 †1.3↓n=3—n=0—n=0—n=0
hard—n=0—n=00.0=n=2 †—n=0—n=0
Median lead timelower is better · n = accepted tasks with lead times · exported measurements
Difficultyclaude-sonnet-5gpt-5.6-terraclaude-opus-4-8claude-fable-5-1gpt-6-astra
easy80 min=n=3—n=0—n=0—n=0—n=0
medium84 min↑n=2 †97 min↓n=3—n=0—n=0—n=0
hard—n=0—n=0119 min=n=2 †—n=0—n=0

↑ best · ↓ worst · = tied or only value · † n < 3, provisional
Grounds compare within a row. Colour is never the only signal.

A task using several models appears in each. Columns do not add up. Exported cohorts are preserved; the build’s shared computation contract is in the design document.

Recorded spend by activity

check
$0.00
other
$32.74
rebase
$0.74
review
$13.29
revise
$4.38
work
$37.75

Runs and hand steps

claude:claude-sonnet-5
10 runs
claude:claude-fable-5-1
2 runs
garden:review
19 runs
claude:claude-opus-4-8
2 runs
codex:gpt-5.6-terra
10 runs
codex:gpt-6-astra
1 runs
garden:edit
21 runs
garden:kickoff
1 runs
codex:gpt-5.6-luna
1 runs
garden:rebase
5 runs
garden:check
29 runs
Hand steps
31 recorded
Manual merges
not supplied
Recorded hand-step kinds
decision_resolved
4
dispatch_paused
1
dispatch_resumed
3
moved
2
suggestion
21

Last 24 hours · 2026-09-05T02:34:47+00:00 → 2026-09-06T02:34:47+00:00

100 merged in this window

$896.16 recorded window spend

All recorded activities in the selected window; price completeness unknown.

Accepted-task cost · separate lifetime measure

$7.15 / accepted task, as exported.

Intended measure: mean lifetime cost of tasks accepted in this window. Exported merged cohort: 100 tasks. Priced denominator and lifetime-cost coverage not supplied; this figure is unverified. It is not window spend divided by merges.

77% first-pass approval

n=102 reviewed tasks (export cohort) · 777 completed runs

Run completions by exported hour bucket: 8, 14, 43, 46, 8, 13, 7, 126, 70, 89, 36, 57, 43, 23, 31, 39, 12, 11, 70, 16, 15.
Accepted-task throughput timestamps not supplied.

12 recorded configuration annotations

2026-09-05T14:36:12+00:00 · config_reloaded · review

2026-09-05T14:59:18+00:00 · config_reloaded · harnesses

2026-09-05T17:11:04+00:00 · config_reloaded · review

2026-09-05T18:50:25+00:00 · config_reloaded · harnesses

2026-09-05T19:45:01+00:00 · config_reloaded · harnesses

2026-09-05T20:37:55+00:00 · config_reloaded · review

2026-09-05T20:39:26+00:00 · config_reloaded · harnesses

2026-09-05T20:45:41+00:00 · config_reloaded · harnesses

2026-09-05T22:32:01+00:00 · config_reloaded · harnesses

2026-09-05T23:07:24+00:00 · config_reloaded · harnesses, review

2026-09-05T23:08:51+00:00 · config_reloaded · review

2026-09-06T02:01:11+00:00 · config_reloaded · harness, review

Scroll to compare models →

Cost per accepted tasklower is better · n = accepted tasks · exported measurements
Difficultyclaude-sonnet-5claude-opus-4-8claude-fable-5-1gpt-5.6-terragpt-6-astra
easy$4.62=n=50—n=0—n=0—n=0—n=0
medium$11.69n=15$8.98n=23$13.36↓n=1 †$5.58↑n=4—n=0
hard—n=0$7.27↑n=2 †$12.20↓n=4—n=0—n=0
First-pass approvalhigher is better · n = reviewed tasks (export cohort) · exported measurements
Difficultyclaude-sonnet-5claude-opus-4-8claude-fable-5-1gpt-5.6-terragpt-6-astra
easy86%=n=49—n=0—n=0—n=0—n=0
medium57%n=1485%↑n=260%↓n=1 †0%↓n=4—n=0
hard—n=0100%↑n=2 †80%↓n=5—n=0—n=0
Work-run costlower is better · n = work runs · exported measurements
Difficultyclaude-sonnet-5claude-opus-4-8claude-fable-5-1gpt-5.6-terragpt-6-astra
easy$2.25=n=73—n=0—n=0—n=0—n=0
medium$2.33n=41$4.56↓n=49$3.83n=3$0.60↑n=12—n=0
hard—n=0$6.24n=2 †$7.31↓n=9—n=0$5.98↑n=1 †
Revise roundslower is better · n = accepted tasks · exported measurements
Difficultyclaude-sonnet-5claude-opus-4-8claude-fable-5-1gpt-5.6-terragpt-6-astra
easy0.3=n=50—n=0—n=0—n=0—n=0
medium1.4↓n=150.2↑n=231.0n=1 †1.2n=4—n=0
hard—n=00.0↑n=2 †0.8↓n=4—n=0—n=0
Median lead timelower is better · n = accepted tasks with lead times · exported measurements
Difficultyclaude-sonnet-5claude-opus-4-8claude-fable-5-1gpt-5.6-terragpt-6-astra
easy47 min=n=50—n=0—n=0—n=0—n=0
medium153 min↓n=1546 min↑n=2355 minn=1 †114 minn=4—n=0
hard—n=0119 min↓n=2 †39 min↑n=4—n=0—n=0

↑ best · ↓ worst · = tied or only value · † n < 3, provisional
Grounds compare within a row. Colour is never the only signal.

A task using several models appears in each. Columns do not add up. Exported cohorts are preserved; the build’s shared computation contract is in the design document.

Recorded spend by activity

check
$0.00
other
$45.66
persona
$119.36
rebase
$20.63
retro
$11.72
review
$153.77
revise
$66.80
work
$478.23

Runs and hand steps

claude:claude-sonnet-5
156 runs
claude:claude-opus-4-8
51 runs
garden:review
200 runs
garden:persona
24 runs
claude:claude-fable-5-1
12 runs
garden:retro
6 runs
codex:gpt-5.6-terra
12 runs
codex:gpt-6-astra
1 runs
garden:edit
21 runs
garden:kickoff
1 runs
garden:compare
2 runs
codex:gpt-5.6-luna
1 runs
garden:rebase
78 runs
garden:check
202 runs
codex:work
2 runs
garden:trial
1 runs
codex:trial
6 runs
garden:work
1 runs
Hand steps
91 recorded
Manual merges
not supplied
Recorded hand-step kinds
budget_set
3
config_override
2
decision_accepted
5
decision_resolved
21
dispatch_paused
5
dispatch_resumed
8
moved
10
resumed
12
suggestion
21
triaged
4

This phase · phase-05 · 2026-09-06T00:01:13+00:00 → 2026-09-06T02:34:47+00:00

10 merged in this window

$88.22 recorded window spend

All recorded activities in the selected window; price completeness unknown.

Accepted-task cost · separate lifetime measure

$4.41 / accepted task, as exported.

Intended measure: mean lifetime cost of tasks accepted in this window. Exported merged cohort: 10 tasks. Priced denominator and lifetime-cost coverage not supplied; this figure is unverified. It is not window spend divided by merges.

42% first-pass approval

n=12 reviewed tasks (export cohort) · 99 completed runs

Run completions by exported day bucket: 99.
Accepted-task throughput timestamps not supplied.

1 recorded configuration annotations

2026-09-06T02:01:11+00:00 · config_reloaded · harness, review

Scroll to compare models →

Cost per accepted tasklower is better · n = accepted tasks · exported measurements
Difficultyclaude-sonnet-5gpt-5.6-terraclaude-opus-4-8claude-fable-5-1gpt-6-astra
easy$2.50=n=3—n=0—n=0—n=0—n=0
medium$5.80↓n=2 †$3.08↑n=3—n=0—n=0—n=0
hard—n=0—n=0$7.27=n=2 †—n=0—n=0
First-pass approvalhigher is better · n = reviewed tasks (export cohort) · exported measurements
Difficultyclaude-sonnet-5gpt-5.6-terraclaude-opus-4-8claude-fable-5-1gpt-6-astra
easy67%=n=3—n=0—n=0—n=0—n=0
medium50%↑n=2 †0%↓n=4—n=0—n=0—n=0
hard—n=0—n=0100%↑n=2 †0%↓n=1 †—n=0
Work-run costlower is better · n = work runs · exported measurements
Difficultyclaude-sonnet-5gpt-5.6-terraclaude-opus-4-8claude-fable-5-1gpt-6-astra
easy$1.94=n=3—n=0—n=0—n=0—n=0
medium$3.32↓n=5$0.72↑n=10—n=0—n=0—n=0
hard—n=0—n=0$6.24n=2 †$11.32↓n=2 †$5.98↑n=1 †
Revise roundslower is better · n = accepted tasks · exported measurements
Difficultyclaude-sonnet-5gpt-5.6-terraclaude-opus-4-8claude-fable-5-1gpt-6-astra
easy0.0=n=3—n=0—n=0—n=0—n=0
medium1.0↑n=2 †1.3↓n=3—n=0—n=0—n=0
hard—n=0—n=00.0=n=2 †—n=0—n=0
Median lead timelower is better · n = accepted tasks with lead times · exported measurements
Difficultyclaude-sonnet-5gpt-5.6-terraclaude-opus-4-8claude-fable-5-1gpt-6-astra
easy80 min=n=3—n=0—n=0—n=0—n=0
medium84 min↑n=2 †97 min↓n=3—n=0—n=0—n=0
hard—n=0—n=0119 min=n=2 †—n=0—n=0

↑ best · ↓ worst · = tied or only value · † n < 3, provisional
Grounds compare within a row. Colour is never the only signal.

A task using several models appears in each. Columns do not add up. Exported cohorts are preserved; the build’s shared computation contract is in the design document.

Recorded spend by activity

check
$0.00
other
$32.74
rebase
$0.74
review
$12.62
revise
$4.38
work
$37.75

Runs and hand steps

claude:claude-sonnet-5
10 runs
claude:claude-fable-5-1
2 runs
garden:review
18 runs
claude:claude-opus-4-8
2 runs
codex:gpt-5.6-terra
10 runs
codex:gpt-6-astra
1 runs
garden:edit
21 runs
garden:kickoff
1 runs
codex:gpt-5.6-luna
1 runs
garden:check
29 runs
garden:rebase
4 runs
Hand steps
31 recorded
Manual merges
not supplied
Recorded hand-step kinds
decision_resolved
4
dispatch_paused
1
dispatch_resumed
3
moved
2
suggestion
21

Inspection sheet · all examples are simulated

Every state has something to say.

The live page uses these treatments in place. This sheet keeps them visible for design review.

Running · beyond typical
in leafwork

Workers on independent hosts

codex · gpt-5.6-terra · medium

Validating the generated context against the project.

Validating the generated context against the project.

typical duration from export · sample n not supplied

$1.24 · recorded at 02:44 UTC

Finishing

Review approved.

Every PR reaches the review queue

2:14 · $0.18 final

Waiting for CI before merge. Keep for eight seconds, then settle into recent activity.

Ⅱ Held

Merge held.

Base branch checks failed.

Other workers continue. This reason stays visible until the hold clears.

View decision · Open in Inbox

Ⅱ Paused

codex is paused.

Account quota reached.

New codex work waits; in-flight clocks keep counting. No retry time reported.

Global variant: Dispatch paused · by operator.

× Failedwilted

Checks failed.

The acceptance check returned an error.

Run stopped at 1:08. The failure stays until a retry or a decision resolves it.

Emptyseed

No tasks yet.

Plan work from the phase goals or create the first task.

Open phase planning · Open task creation

Next: no approved work
Phase: no tasks yet
Last period: no recorded outcomes

Drafts awaiting approval

Work is ready to consider.

Review existing drafts before workers can start.

Review drafts

Quiet

Nothing running.

Waiting for an approved task.

Next tick · 02:46 UTC

Phase and past outcomes stay visible. There is no idle animation.

Disconnected

Updates disconnected.

Last received at 02:44 UTC.

Retain the last snapshot. Clocks show elapsed since start; they cannot prove a worker is still alive.

Unavailable

State not supplied.

Queue unavailable · typical not established · progress not mapped.

Costs and outcomes show —. Missing history is not an empty garden.

Now 2 · design target for /now2 · recorded snapshot + simulated atlas · full specification in docs/design/now-2.md