Remove the depth-2 cap in walk_meta_inputs so every fetched input
at every depth is surfaced, not just two levels (issue #275). The
uncapped walk needs a guard: a visited-node set makes it a spanning
tree — each fetched node walked once, at its shallowest path — so
shared subtrees don't re-walk and a cycle can't recurse forever.
A two-pass walk (claim a node's direct inputs before descending)
keeps shallow inputs at a shallow path.
Frontend: renderMetaInputs indents each row by its slash-path depth
and shows the leaf segment (full path on hover), plus a select-all /
select-none control so a long input list isn't ticked box by box.
the stream-json result event carries modelUsage.<model>.contextWindow
which is the actual per-inference active window the model enforces.
for claude-sonnet-4-6 this is 200k even though the full prompt cache
can hold millions of tokens via accumulated cache reads.
with the nix-configured sonnet = 1000000 the proactive compact watermark
sat at 750k and was never reached. agents grew context until prompt_too_long
at ~170k — reactive compact, no checkpoint turn.
changes:
- bus gains api_context_window field seeded from modelUsage.*.contextWindow
in each turn's result event. authoritative; falls back to env var, then 200k.
- new effective_context_window(bus) helper used by both watermark functions
- compact_watermark (75%) and auto_reset_watermark (50%) call effective_context_window
- context_tokens() docstring clarified: all three token fields (input +
cache_read + cache_creation) count against the per-inference contextWindow
limit. the large cache_read values seen in the result event are cumulative
across all inferences in a turn, not per-inference.
- /api/state context_window_tokens now reflects the calibrated window
closes#129
extract per-agent forge logic from ensure_all() into sync_agent()
so both the startup sweep and rebuild_agent call identical code.
rebuild now runs: ensure_user_for + ensure_config_repo + push_config
+ meta_read_access + ensure_meta_remote — same as the boot sweep.
missing tokens and drift in any forge state are fixed by rebuild,
not just hive reboot.
if forge_after_first_spawn fails transiently on first spawn the
token is missing. rebuild_agent now calls ensure_user_for so
a manual rebuild (or the startup auto-update scan) recovers
the missing token — no full hive reboot needed.
post_meta_update returns 200 immediately and runs the nix flake
update + agent-rebuild ripple in a background task, so the META
INPUTS panel looked idle for the whole multi-minute window (#259).
Track in-flight runs with a Coordinator atomic counter, exposed via
an RAII MetaUpdateGuard held across run_meta_update. Surface it as
the meta_update_running snapshot field plus a MetaUpdateRunning SSE
event (flipped only when the count crosses 0, so concurrent runs
flip the flag once). The panel shows a pulsing in-progress banner
and disables the update button while a run is active.
