refactor(server): fetch boot context from ContexGin HTTP API

[dimakis]

Summary

• Replace inline import('contexgin') library call with HTTP fetch to running ContexGin server (GET /api/agents/:name/context)
• The old approach used a hardcoded 8k token budget fallback and read the agent recipe from the worktree path (often stale/missing). The server already compiles correctly with the recipe's 12k budget.
• Extract fetchBootContext() as a testable function, reducing ~140 lines of library import + recipe parsing + response mapping to a single HTTP call

Root cause

chat.ts imported contexgin as a Node library and called the legacy compile() function with DEFAULT_TOKEN_BUDGET = 8000. It attempted to read tokenBudget from .agents/mitzo-conversational.yaml in cwd (the worktree), which often didn't have the file or had a stale version. Meanwhile, the running ContexGin server at CONTEXGIN_URL was already compiling correctly with 12k. Mitzo was bypassing its own infrastructure.

Test plan

• 7 new tests: success path, unreachable server, non-200 response, missing boot field, empty sources, non-string source filtering, env var URL
• Existing contexgin-smoke, context-baseline, and prompt-compare tests still pass (21/21)
• Full test suite: no new failures (2 pre-existing on main)
• Manual: start a Mitzo session, verify boot context shows >8k budget in source explorer

🤖 Generated with Claude Code

[dimakis]

Centaur Review

Found 5 issue(s) (2 warning).

server/chat.ts

Clean refactor from library import to HTTP API. Main concern is a missing .catch() on the fire-and-forget promise (unhandled rejection risk), and a UI regression where the included/trimmed section detail and budget-vs-actual distinction are silently lost from BootContextPill.

• 🟡 bugs (L835): Missing .catch() on fire-and-forget promise chain. If transport.send() throws after the async gap (e.g., transport closed between fetch and send), this produces an unhandled promise rejection. The adjacent capturePromptComparison call correctly uses .catch(() => {}). Fix: fetchBootContext('mitzo-conversational').then((msg) => send(transport, msg)).catch(() => {}); [fixable]
• 🟡 regressions (L160): included and trimmed are now always empty arrays. The old code populated these from compiled.included/compiled.trimmed with section-level detail (source, heading, token count, content). The frontend's BootContextPill renders expandable 'Included (N)' and 'N sections trimmed' sections from these fields — those UI sections will never appear now. If intentional (ContexGin manages compilation internally), these fields on BootContextMessage are effectively dead code and should be removed from the interface to avoid confusion.
• 🔵 regressions (L160): tokenBudget is now always equal to tokenCount (both set to bootTokens). The old code distinguished configured budget (from agent recipe YAML, default 8000) from actual usage. The frontend displays these as 'X.Xk / X.Xk' (count / budget) which is now always identical, making the budget half of the label redundant. In the fallback case, budget changed from 8000 to 0, so the pill shows '0 / 0' instead of '0 / 8.0k'.

server/__tests__/boot-context.test.ts

Clean refactor from library import to HTTP API. Main concern is a missing .catch() on the fire-and-forget promise (unhandled rejection risk), and a UI regression where the included/trimmed section detail and budget-vs-actual distinction are silently lost from BootContextPill.

• 🔵 missing_tests: Missing test for malformed JSON response — when res.json() throws (e.g., body is not valid JSON despite ok: true), the outer catch block should return FALLBACK_BOOT_CONTEXT. This edge case is handled correctly by the implementation but not verified by tests. [fixable]
• 🔵 unsafe_assumptions (L167): Env var cleanup in the last test is not wrapped in try/finally. If an assertion fails before line 180, process.env.CONTEXGIN_URL leaks its test value into subsequent tests. Use try/finally or Vitest's vi.stubEnv() for safer cleanup. [fixable]

[dimakis]

🟡 bugs: Missing .catch() on fire-and-forget promise chain. If transport.send() throws after the async gap (e.g., transport closed between fetch and send), this produces an unhandled promise rejection. The adjacent capturePromptComparison call correctly uses .catch(() => {}). Fix: fetchBootContext('mitzo-conversational').then((msg) => send(transport, msg)).catch(() => {}); [fixable]

[dimakis]

🟡 regressions: included and trimmed are now always empty arrays. The old code populated these from compiled.included/compiled.trimmed with section-level detail (source, heading, token count, content). The frontend's BootContextPill renders expandable 'Included (N)' and 'N sections trimmed' sections from these fields — those UI sections will never appear now. If intentional (ContexGin manages compilation internally), these fields on BootContextMessage are effectively dead code and should be removed from the interface to avoid confusion.

[dimakis]

🔵 regressions: tokenBudget is now always equal to tokenCount (both set to bootTokens). The old code distinguished configured budget (from agent recipe YAML, default 8000) from actual usage. The frontend displays these as 'X.Xk / X.Xk' (count / budget) which is now always identical, making the budget half of the label redundant. In the fallback case, budget changed from 8000 to 0, so the pill shows '0 / 0' instead of '0 / 8.0k'.

[dimakis]

🔵 unsafe_assumptions: Env var cleanup in the last test is not wrapped in try/finally. If an assertion fails before line 180, process.env.CONTEXGIN_URL leaks its test value into subsequent tests. Use try/finally or Vitest's vi.stubEnv() for safer cleanup. [fixable]