Flexy

Toggle OpenAI Flex in Pi, check whether your last call used it, and see estimated savings.

Install

pi install npm:@dowhilegeek/pi-flexy

Start Pi or run /reload in an existing session. Installation works even if Flexy isn't listed in Pi's package gallery.

Requires Node ≥22.19, Pi ≥0.84.4 and <0.85.0, and an OpenAI API key. ChatGPT/Codex subscriptions aren't supported.

Use

In Pi, use /login to add your OpenAI API key, then /model to select a Flex-supported model under openai.

Start Pi with Flex enabled before first prompt:

pi --flex

Flexy retries transient, zero-output Flex failures twice by default. Set a different retry budget at startup if needed:

pi --flex --flex-retries 4

--flex-retries 4 means four retries after the first attempt, for five total stream attempts. Accepted values are 0 through 10.

You can also run /flex on in an existing session. Send a prompt, then run /flex audit after the reply.

Command	What it does
/flex on	Use Flex for subsequent calls
/flex off	Return to standard processing
/flex toggle	Switch Flex on or off
/flex status	Show current mode and last-call summary
/flex audit	Check whether the last call was sent and served as Flex
/flex savings	Show estimated savings totals for the current session branch
/flex savings all	Roll up all locally saved sessions and all branches
/flex savings all json	Full rollup with every session/model/day group and scan warnings
/flex history	Show recent calls
/flex retries	Show retry budget
/flex retries N	Save global retry budget, from 0–10
/flex fallback	Show standard-tier fallback preference
/flex fallback on|off	Save global fallback preference (default off)
/flex help	Show all options

The footer shows 💪 flex:on or 💪 flex:off. New sessions start with Flex off unless launched with --flex.

Both --flex and /flex on affect only models using Pi's native openai provider and openai-responses API. For a Codex subscription or any other provider, Flexy renders no status line at all and leaves requests untouched; /flex status and /flex audit still name the active model and why it is out of scope.

Savings across sessions

/flex savings shows current-branch totals and coverage, without a last-call section. /flex savings json retains its existing machine-readable schema, including lastCall, for compatibility. Use /flex audit json for a machine-readable last-call audit. The old --all and --json spellings still work as aliases.

Use /flex savings all for lifetime estimates across locally retained sessions. It scans ~/.pi/agent/sessions (or $PI_CODING_AGENT_DIR/sessions), the current session's storage directory, and $PI_CODING_AGENT_SESSION_DIR when set. It includes inactive branches and current in-memory entries, not just the last 50 audits. Archives in unrelated custom directories and deleted histories cannot be discovered automatically.

The report shows:

• Estimated spend, equivalent standard-price cost for the same token/cache usage, and savings.
• Breakdown by model (dollars and percentage saved), top 10 sessions by savings, and latest 10 UTC call-start dates. Model percentages are savings divided by equivalent standard-price cost across that model's included Flex and standard calls—not an average of per-call percentages. all json includes every group; argument order does not matter.
• Included/excluded calls, unaudited histories, deduplicated copies, and scan warnings.

Repeated audit snapshots and copied fork histories do not count as new spending. Calls deduplicate by audit ID and provider/API response ID. A call belongs to its earliest retained session copy for the session breakdown. Contradictory evidence or ambiguous identities are excluded rather than inflating totals. Historical audits can use matching saved Pi costs; today's model catalog prices are never substituted for missing historical prices. Standard-tier fallback contributes zero Flex savings.

Scanning is read-only: no session migration, file writes, API requests, or automatic caching. Each file is read only up to its size when opened. Unterminated tails are deferred, corrupt/oversized records are skipped with warnings, and symlinks are not followed. Only accounting metadata is retained by the scanner; transcript content is not included in the report or sent to a model. Empty or unpriceable histories report unknown, not zero.

These remain estimates for completed responses, not invoice totals. Failed-attempt charges, unaudited tool/summary work, taxes, and account-specific adjustments are excluded. Active sessions can change during the scan; run the command again to refresh.

Retries and standard-tier fallback

Retryable errors can arrive inside an HTTP 200 stream, not just as HTTP errors. Flexy handles retries inside the original provider call, so queued steering does not enter the retried request. Error classification uses Pi's transient-error classifier; it does not imply a particular HTTP status or prove a Flex capacity rejection.

By default, Flexy allows two Flex retries after the initial attempt, with fallback off. Backoff starts at two seconds, doubles, and caps at 30 seconds. Text, reasoning, or tool-call output prevents automatic replay. Cancellation and non-retryable errors do not trigger fallback.

Enable one standard-tier attempt after the Flex budget is exhausted:

/flex retries 2
/flex fallback on

That sequence permits Flex → Flex → Flex → default. For an immediate switch after the initial Flex failure, use /flex retries 0 with fallback enabled: Flex → default.

Fallback uses standard pricing. It changes only the failed provider call, not your session's Flex mode. The next new model call starts on Flex again. Fallback gets exactly one attempt; if it fails, Flexy stops rather than returning to Flex or allowing Pi to add more retries. Managed Flex calls disable SDK-level retries so HTTP retries cannot multiply this budget. Standard-mode calls keep their normal Pi behavior.

Retries use a detached snapshot of the final post-hook JSON payload. Fallback changes only service_tier to default; it does not rerun payload hooks or include queued input. /flex audit records each HTTP attempt, the retry budget, whether fallback was used, and bounded provider rejection details (code, type, param, reason, and message) before a retry can hide them. Adapter-level stream errors are retained per attempt when no structured provider error exists. Messages are capped at 32,768 characters and other error fields at 1,024 characters, with truncation reported. A successful default-tier response contributes no Flex savings.

Session activity

Retries and fallback appear live in the Pi transcript, not just in transient notifications:

[2026-09-09 20:10:11.123 UTC] Flexy: Flex request failed. Retry 1/2 scheduled in 2s (still Flex pricing).
[2026-09-09 20:10:13.125 UTC] Flexy: Retrying failed request on Flex (1/2).
...
[2026-09-09 20:10:25.400 UTC] Flexy: Falling back to non-Flex after 3 failed Flex attempt(s).
Requesting default tier at standard pricing for this call only. Session Flex mode unchanged.

The fallback notice uses warning coloring and is recorded when the default-tier request reaches the HTTP transport. It reports requested pricing, not a billing confirmation; /flex audit shows the response-tier evidence. Cancelling before that handoff does not produce a fallback marker.

Entries retain their original UTC timestamps across resume and reload. Expand an entry to see its audit call ID. The footer also shows retry/backoff or the current standard-tier fallback while active. These are custom session entries, never model messages: they do not steer the agent, trigger extra turns, or enter prompts/compaction context. Activity entries store no prompts, generated output, or provider errors; bounded provider rejection fields live only in the corresponding audit entry and /flex audit output. Print mode writes the same activity text to stderr; JSON/RPC clients receive native entry_appended events.

Preferences across sessions

/flex retries N and /flex fallback on|off save preferences in ~/.pi/agent/flexy.json (or $PI_CODING_AGENT_DIR/flexy.json). Existing sessions read preferences before each new managed call. In-flight calls keep the policy they started with. Branch navigation does not rewind these preferences; older branch-local retry settings are superseded by global preferences. Flex on/off mode remains branch-local.

Startup flags also save these preferences:

pi --flex --flex-retries 0 --flex-fallback on

Omit those preference flags to keep saved values; --flex-fallback off disables fallback. Invalid configuration produces a warning and uses two retries with fallback off, without rewriting the file. Failed preference writes are reported instead of silently changing the active policy.

Flex can still be slow or unavailable. Fallback is never enabled implicitly. Savings are estimates, not billing records.

Inspired by pi-flex-processing.