chore(agents): strip [1m] long-context suffix from every default model #394

Merged

charles merged 1 commit from chore/strip-1m-context-defaults into main

2026-04-26 19:51:23 +00:00

claude-desktop commented

2026-04-26 18:51:59 +00:00

Collaborator

Summary

1M context tokens always meter against the Anthropic extra-usage rail, not against the Pro Max included weekly quota. The €200/mo extra-usage cap was hit (€200,63 / €200) on 2026-04-26, blocking every reviewer / boss / designer / foreman dispatch with API Error: Extra usage is required for 1M context · enable extra usage at claude.ai/settings/usage even though the Pro Max Sonnet weekly bucket sat at 5%.

This drops the [1m] suffix from:

config/agents.json — boss, reviewer, designer, foreman defaults
apps/server/src/domain/agent/foreman.ts — fallback when the request omits a model
apps/server/src/background/worker.ts — stale "bump model to claude-opus-4-7[1m]" advice in the prompt-too-long hint for designer (split the task instead)
apps/web/e2e/planner.spec.ts — fixture aligned with the new default

Validation

Round-tripped on PR #393 (the issues.unassigned fix): reviewer-2 ran on plain claude-sonnet-4-6 and APPROVED with substantive feedback (two doc nits flagged non-blocking). 200k headroom is sufficient for single-PR review even with the forgejo-mcp tool surface.
Typecheck clean
[1m]-fixture tests still pass (parsing/round-trip behaviour preserved)

Out of scope

The [1m]-stripping logic in token-economy.ts stays as-is — operators can still override individual instances via the agents DB when a specific job genuinely needs 1M (in which case extra-usage spend is intentional and signed off).

🤖 Generated with Claude Code

## Summary 1M context tokens always meter against the Anthropic **extra-usage** rail, not against the Pro Max included weekly quota. The €200/mo extra-usage cap was hit (€200,63 / €200) on 2026-04-26, blocking every reviewer / boss / designer / foreman dispatch with `API Error: Extra usage is required for 1M context · enable extra usage at claude.ai/settings/usage` even though the Pro Max Sonnet weekly bucket sat at 5%. This drops the `[1m]` suffix from: - `config/agents.json` — boss, reviewer, designer, foreman defaults - `apps/server/src/domain/agent/foreman.ts` — fallback when the request omits a model - `apps/server/src/background/worker.ts` — stale "bump model to claude-opus-4-7[1m]" advice in the prompt-too-long hint for designer (split the task instead) - `apps/web/e2e/planner.spec.ts` — fixture aligned with the new default ## Validation - Round-tripped on PR #393 (the issues.unassigned fix): reviewer-2 ran on plain `claude-sonnet-4-6` and APPROVED with substantive feedback (two doc nits flagged non-blocking). 200k headroom is sufficient for single-PR review even with the forgejo-mcp tool surface. - Typecheck clean - `[1m]`-fixture tests still pass (parsing/round-trip behaviour preserved) ## Out of scope - The `[1m]`-stripping logic in `token-economy.ts` stays as-is — operators can still override individual instances via the agents DB when a specific job genuinely needs 1M (in which case extra-usage spend is intentional and signed off). 🤖 Generated with [Claude Code](https://claude.com/claude-code)

claude-desktop added 1 commit

2026-04-26 18:51:59 +00:00

chore(agents): strip [1m] long-context suffix from every default model

qa / qa (pull_request) Successful in 6m56s

Details

qa / dockerfile (pull_request) Successful in 11s

Details

1a4bde2989

1M context tokens always meter against the Anthropic extra-usage rail,
not against the Pro Max included weekly quota. The €200/mo extra-usage
cap was hit (€200,63 / €200) on 2026-04-26, blocking every reviewer /
boss / designer / foreman dispatch with `API Error: Extra usage is
required for 1M context · enable extra usage at claude.ai/settings/usage`
even though the Pro Max Sonnet weekly bucket sat at 5%.

Drop the `[1m]` suffix from:
- `config/agents.json` — boss, reviewer, designer, foreman defaults
- `apps/server/src/domain/agent/foreman.ts` — fallback when the
  request omits a model
- `apps/server/src/background/worker.ts` — stale "bump model to
  claude-opus-4-7[1m]" advice in the prompt-too-long hint for designer
  (split the task instead)
- `apps/web/e2e/planner.spec.ts` — fixture aligned with the new default

Round-tripped on PR #393 (the issues.unassigned fix): reviewer-2 ran
on plain `claude-sonnet-4-6` and APPROVED with substantive feedback,
so 200k headroom is sufficient for single-PR review even with the
forgejo-mcp tool surface.

The `[1m]`-stripping logic in `token-economy.ts` stays — operators can
still override individual instances via the agents DB when a specific
job genuinely needs 1M (in which case extra-usage spend is intentional).

charles requested review from reviewer

2026-04-26 19:48:27 +00:00

reviewer approved these changes

2026-04-26 19:49:34 +00:00

reviewer left a comment

All four sites correctly stripped. The worker.ts hint improvement (drop the "bump to [1m]" advice, keep the "split the task" remediation) is a nice bonus. CI green.

All four sites correctly stripped. The `worker.ts` hint improvement (drop the "bump to [1m]" advice, keep the "split the task" remediation) is a nice bonus. CI green.

charles merged commit 445c8e2a60 into main

2026-04-26 19:51:23 +00:00

charles deleted branch chore/strip-1m-context-defaults

2026-04-26 19:51:24 +00:00

charles referenced this pull request from a commit

2026-04-26 19:51:24 +00:00

chore(agents): strip [1m] long-context suffix from every default model (#394)